Répertoire des ressources
Trouvez ce dont vous avez besoin pour votre flux de travail.
Guides: techniques, setup, and troubleshootingRecherchez Sweet Tea, Pipes, des flux de travail externes, des modèles, des packs de nœuds et des guides dans un catalogue unique lié aux preuves.
Commencez plutôt avec un flux de travail ↗Résultats de la recherche
Toutes les ressources
20 affiché of 42 663
Wan 2.1 transparent animated WebP
Generate a transparent animated WebP from a text description using Wan 2.1 and its RGBA adaptation. Color and alpha are decoded from the same generated video latents and combined before encoding. No uploaded image is required. For a first run, replace the example prompt with your subject, movement, framing, and visual style, and request a transparent background. Keep 832 × 480 pixels, 33 frames, batch size 1, and 16 playback FPS for approximately 2.06 seconds. Width and height use multiples of 16; frame counts follow 1, 5, 9, and so on. Larger dimensions and more frames increase generation work. FPS changes playback speed and duration without generating more frames. Advanced Sampling controls include Seed, Steps, CFG, sampler, scheduler, and denoising amount. Seed -1 randomizes each run; use a fixed seed to compare settings. The default configuration uses 20 steps, CFG 6, uni_pc, simple scheduling, and full denoising. The negative prompt lists unwanted content and artifacts; its supplied default is in Chinese. Lightx2v is optional and bypassed by default. To use the documented accelerated configuration, enable that stage and manually set Steps to 4 and CFG to 1. Its strength has no effect while bypassed. Keep RGBA adaptation strength 1 and sampling shift 5 for a first run. RGBA adaptation strength changes the model, not output opacity. WebP output exposes playback FPS and encoding controls. Lossless encoding is enabled by default. Disable it for lossy encoding, where Quality controls detail retention; in lossless mode Quality controls compression effort. Keep Compression method at default initially. Keep batch size 1 for one animation: this selected output encodes the resulting image batch into animated WebP and does not provide separate video outputs. The result is animated WebP rather than MP4. Use a viewer or compositing tool that supports animated WebP and alpha. Generated subject boundaries and fine details may need review before compositing.
Ideogram 4 image generator
Generate images from scene prompts using the fixed Ideogram 4 conditional and unconditional model pair. Enter your completed scene description in Positive prompt; structured JSON is recommended for compositions containing text and positioned objects. No uploaded media is required. For a first run, keep the supplied monochrome knight-and-horse magazine-cover prompt, square aspect ratio, 1 megapixel output, Default sampling preset and Image count 1, then generate. The example mentions image_0f5a20.png, but there is no reference-image input and that file is not supplied to the model. Replace that mention with a written typography description when adapting the example. Technical configuration text inside the prompt is requested lettering in the image; it does not change the actual sampler, steps or seed. Output size controls both the latent canvas and scheduler dimensions. Dimensions round upward to multiples of 16, with a minimum of 256 pixels per side. Changing the canvas aspect ratio does not rewrite the prompt's JSON aspect_ratio or bounding boxes; adjust those manually to suit your composition. Optional boxes use [y1,x1,y2,x2] coordinates normalized to 0–1000. Sampling preset selects coordinated scheduler settings: Default uses 20 steps, mu 0 and std 1.75; Quality uses 48 steps, mu 0 and std 1.5; Turbo uses 12 steps, mu 0.5 and std 1.75. These names identify sampling configurations rather than guaranteed output quality. Advanced controls expose guidance, a guidance override interval, seed, sampler and image count. Main CFG defaults to 7; Override CFG defaults to 3 between progress 0.7 and 1.0. Keep Override start below Override end. Set the override equal to main CFG to remove their guidance difference. Seed -1 selects fresh noise each run; a fixed seed allows comparisons with the same noise initialization. Euler is the source sampler default. Larger image counts and resolutions require more memory and computation. The workflow uses ideogram4_int8_convrot.safetensors and ideogram4_unconditional_int8_convrot.safetensors in diffusion_models, qwen3vl_8b_fp8_scaled.safetensors in text_encoders, and flux2-vae.safetensors in vae. These fixed models must be available in the execution environment. The supplied text-encoder installation source is https://huggingface.co/Comfy-Org/Qwen3-VL/resolve/main/text_encoders/qwen3vl_8b_fp8_scaled.safetensors. There is no independent negative prompt: negative conditioning is zeroed from positive conditioning and passed to the unconditional model. The source's separate external-LLM instruction template is outside this image output. This Pipe does not call an LLM or automatically turn an idea into JSON; paste your finished prompt directly into Positive prompt.
SeedVR2 image upscale
Create an enlarged, restored image with SeedVR2 and optional source color matching. Upload the image to restore, leave Scale factor at 4 and Color matching at none, then run. The default multiplies both dimensions by 4, producing approximately 16 times the source pixel count before restoration. No text prompt is required: the uploaded image supplies conditioning. The model is fixed to SeedVR2 3B int8 with its matching VAE. Scale factor controls the initial resize: values above 1 enlarge, 1 retains the source dimensions and values below 1 reduce them. SeedVR2 restoration still runs in each case. Resampling method controls interpolation before restoration; lanczos is the source default for enlargement, area supports reduction and nearest-exact retains hard pixel boundaries during resizing. Restoration may subsequently change those details. Color matching uses the resized source as its reference. Choose none to skip color transfer, lab for CIELAB color transfer, wavelet to transfer broad color while retaining restored high-frequency detail, or adain to match per-channel mean and standard deviation. Advanced restoration settings include Seed, Denoise, Steps, Conditioning guidance, Sampler and Scheduler. Seed -1 randomizes each run; use a fixed seed to keep the sampling noise consistent when comparing settings. For a first run, retain denoise 1, one step, guidance 1, Euler sampling and the simple schedule. Other sampling configurations are experimental for this source setup. Advanced input and output tiling independently control spatial encoding and decoding. Tile size and overlap are in pixels; keep each overlap below its corresponding tile size. Larger output dimensions increase the amount of image data processed. Restoration can reinterpret fine details rather than recover exact missing information. Uploaded image and alpha-mask outputs are joined before resizing and preprocessing, but final output transparency is not established; do not rely on this tool to preserve transparency.
DALL·E 2 image generation
Generate one or more square images from a text prompt with DALL·E 2. For a first run, replace the example prompt with your subject, composition, lighting and visual style, keep Output size at 1024×1024 and Image count at 1, then generate. Choose 256×256 or 512×512 for smaller images, or request up to eight images using the same prompt and size. Results are returned as an image batch for preview; this workflow does not permanently save them. Generation sends your prompt to OpenAI and uses a paid provider call. Only text-to-image generation is connected: reference-image editing and masked inpainting are unavailable in this Pipe. The source backend does not implement seed, so a fixed seed does not provide reproducible results.
First and last frame to video · LTX-2.5
Generate one video with generated audio, guided by a starting image, an ending image and a motion prompt. Both images are required endpoint guides; they do not guarantee exact copies of the uploaded frames. For a first run, upload both images and describe the action connecting them, camera movement and desired sound. Keep 1280 × 720, 5 seconds, 24 FPS, endpoint strengths of 0.7 and prompt enhancement enabled. Both uploads are center-cropped to the shared dimensions. Separate advanced image-fit controls can stretch either image instead. Uploaded alpha masks are not connected to generation. Prompt enhancement uses your text and the resized, compressed first image; it does not inspect the last image. Turn enhancement off to use your written prompt directly. Within enhancement tuning, sampling on/off selects sampled or unsampled text expansion independently of the enhancement toggle. Temperature, seed and token sampling filters apply to sampled expansion. The length limit and built-in template are also available. Thinking mode is conditional on model support, which the supplied evidence does not establish for the configured model. Use positive whole-second durations and integer frame rates from 1 to 120. The graph requests duration × FPS + 1 frames: keep duration × FPS divisible by 8 and the resulting count at or below 1000. Playback can be about one frame longer than requested. Duration is entered manually. Use positive image dimensions of at least 64 pixels and keep the supplied dimensions for the first run. Larger dimensions or longer clips increase generation and decoding work. Adjust the starting and ending guide strengths independently to change each image's influence. Video and audio text-guidance scales are separate and default to 1; negative conditioning may have little or no effect at those defaults. The negative prompt starts empty so you can add exclusions suited to your own scene. Audio is generated jointly with the video and included in the returned clip; this interface has no audio upload or silent-output mode. Generation noise and sampled prompt enhancement have separate seeds. Use -1 for fresh randomization. For comparisons with sampled enhancement enabled, fix both seeds. Advanced controls also expose image compression, ancestral noise and the comma-separated distilled noise schedule. Keep their source defaults for the first run. Video and audio batch counts must match; keep both at 1 for the supported single-video output. Separate clips from larger batches are not established by this graph. This is a fixed LTX-2.5 distilled model recipe. Install ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors in diffusion_models; gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors in text_encoders; and ltx-2.5-video-vae-bf16.safetensors plus ltx-2.5-audio-vae-bf16.safetensors in vae. Their source is https://huggingface.co/Lightricks/LTX-2.5. Prompt enhancement additionally uses gemma4_e2b_it_bf16.safetensors in text_encoders from https://huggingface.co/Comfy-Org/gemma-4. Host authorization may be required to obtain model files.
HunyuanVideo text to video
Generate a silent video from a scene and motion prompt using HunyuanVideo. Describe the subject, setting, visual style, action and camera movement. No uploaded image or video is required. For a first run, replace or edit the example prompt and keep 848 × 480 pixels, 73 frames, 24 FPS and 20 sampling steps. This produces approximately three seconds of video. Use dimensions divisible by 16 and frame counts of the form 4n + 1, such as 49, 73 or 97. Larger dimensions and more frames increase computation and memory requirements. Frames determines how many frames are generated. FPS changes playback timing without generating additional frames: duration is approximately Frames divided by FPS. Keep 8-bit initially; 10-bit offers smoother gradients but requires compatible encoding and playback support. Advanced controls include prompt guidance, seed, sampling steps, sampler, scheduler and sampling shift. Seed -1 chooses fresh starting noise each run; use a fixed seed when comparing settings. Start with the supplied generation settings. The sampling shift modifies the model used by the guider, while the scheduler receives the original model, following the source graph. The sampling schedule fraction defaults to 1 for generation from an empty latent; lowering it shortens the schedule and does not preserve an uploaded image. Keep latent batch size at 1 for a first run. Increasing it samples multiple latent videos, but the workflow assembles decoded frames into one video output rather than returning separate clips. Tiled decoding exposes spatial tile size and overlap, plus temporal frame chunk size and overlap. Start with the supplied settings and keep each overlap smaller than its corresponding tile or chunk size. Video encoding offers automatic, MP4, MKV or WebM container selection and automatic, H.264 or AV1 codec selection. Keep format and codec on auto initially. Explicit selections must be compatible with each other, the bit depth and the intended player. CRF controls encoding compression separately from generation; -1 uses the output node's default, and explicit values depend on the selected codec. The workflow uses a fixed HunyuanVideo model, matching text encoders and a video VAE. It has one positive text prompt, no negative-prompt branch, no audio input and no image conditioning. The selected result is one silent video.
Gemini reference images to video
Generate one 720p, 24 FPS video with audio from two reference images and a scene or story prompt using Gemini Omni Flash. For a first run, upload both required images: a primary character, scene or style reference as Reference image 1, and a storyboard or complementary reference as Reference image 2. Replace the example prompt with your own action, visual style, camera behavior and desired sound. Explain how each image should guide the result. For a storyboard, describe panel reading order and the action to animate. Use 9:16 for portrait or 16:9 for landscape, and a requested duration from 3s through 10s. The defaults request a 3-second portrait clip. These values are appended to the prompt as provider instructions; avoid conflicting duration or aspect-ratio instructions in your main prompt. Both images are general references, without guaranteed start-frame or end-frame roles. Only their image outputs are supplied; uploaded alpha masks are not passed to generation. Generation tuning contains Temperature, Top P and Seed. Temperature controls sampling randomness; Top P controls the cumulative probability threshold for nucleus sampling. Keep their defaults for a first run. Seed -1 randomizes the re-run seed each run. Seed affects re-running and caching; a fixed seed does not reproduce the same video. Video output offers optional file type, codec and encoding-quality settings. Leave file type and codec on auto and CRF at -1 for the first run. These settings affect output encoding, rather than the provider's creative generation. Generation sends your prompt and reference images to a hosted Gemini service, requires configured provider access and uses provider billing.
Product placement photo
Generate a still product-placement image from three product photos and a composition reference. Upload all four required images: products occupy images 1–3, and the scene reference is image 4. In Placement prompt, describe how those products should appear and which aspects of the reference to retain or change. For a first run, keep the default model, auto aspect ratio and 1K resolution. The default prompt assumes products held in a hand in a fanned arrangement; revise those instructions if your reference depicts another composition. Auto follows an input image's aspect ratio, but the source does not specify which of the four takes precedence. Choose an explicit aspect ratio when framing matters. Model selects Nano Banana 2 (Gemini 3.1 Flash Image) or Nano Banana 2 Lite. Both use all four references and the same aspect-ratio and thinking-level controls. For Image resolution, enter 1K, 2K or 4K with the full model; Lite supports only 1K, so set 1K before switching to Lite. Thinking level offers MINIMAL or HIGH; HIGH requests additional model reasoning. Only the final image is returned, not text or thinking images. Seed -1 randomizes each run. A fixed seed requests best-effort repeatability, without guaranteeing identical output. Temperature controls generation randomness from 0 to 2; nucleus sampling controls the sampling threshold from 0 to 1. Lower values favor more focused choices. Keep their supplied defaults initially. Image generation instructions optionally change the generator's foundational instructions; use Placement prompt for ordinary scene changes. Inspect product identity, packaging and label text before using the result: the prompt requests fidelity, but exact reproduction is not guaranteed. Uploaded alpha masks are not connected, so uploaded transparency is not preserved through a separate mask input. Generation uses a provider API and requires its runtime access. This tool returns a still image. The separate product-placement video tool also generates the composition, writes motion direction and animates it.
Z-Image Turbo with PiD 4× upscale
Generate a 4096 × 4096 square image from text with Z-Image Turbo and PiD, and receive the 1024 × 1024 base preview alongside it. The same prompt guides both stages; PiD uses the generated base latent as conditioning for its final pixel-space image. For a first run, replace the example prompt with your subject, composition, lighting and style, then keep the supplied defaults: eight base steps and four PiD steps, guidance of 1 for both stages, res_multistep base sampling and LCM PiD sampling with simple schedules. No media upload is required or supported. The base and PiD seeds control separate stages. Use -1 to randomize each run, or fix both seeds when comparing settings. Changes to base sampling affect both the base preview and the latent that conditions PiD. PiD sampling controls affect the final image. Advanced settings expose each stage's sampler, guidance and schedule, plus base sampling shift, denoising fractions and PiD noise initialization. Keep both denoising fractions at 1 and Add PiD noise enabled for the supplied setup. Disabling PiD noise initialization removes the PiD seed's effect on initial noise. Latent degradation is separate from denoising fraction: zero denotes a clean base latent, while higher values tell PiD to handle corrupted latent outputs. The model pair, square dimensions and single-image batches are fixed for this 1024-to-4096 setup. Both stages run to produce the final result; there is no base-only or uploaded-image upscale mode. There is no editable negative prompt: both stages use zeroed negative conditioning. Alternative sampling settings change the generation behavior without a guaranteed quality improvement.
Magazine cover
Create a magazine-cover image from a reference photo or a written description. For a first run, upload a photo, replace the masthead and cover lines in Cover prompt, and keep the portrait 3:4 aspect ratio and 2K resolution. To generate from text alone, disable the Source image stage and describe the subject, background and typography in Cover prompt. Choose an explicit aspect ratio for text-only generation. The default recipe requests flat, minimalist editorial typography: a large pale-yellow masthead, small white cover lines, asymmetric placement and clean negative space. Replace the example wording with your own text. For a street-poster border collage, request cropped lettering, rotated microtext and symbols around the edges while keeping the center and subjects clear. For a foreground-subject variation, explicitly ask for the subjects to appear in front of the typography. Apply these recipes by editing Cover prompt; they are generative instructions rather than separate stages or editable layers. Model selects between gemini-3-pro-image-preview and Nano Banana 2 (Gemini 3.1 Flash Image). Aspect ratio sets the cover's width-to-height proportions; auto follows a supplied reference image. Resolution offers 1K, 2K and 4K, with the provider's native upscaler used for 2K and 4K. Seed -1 randomizes each run; a fixed seed requests best-effort consistency without guaranteeing identical results. The result is a generated image rather than editable text or layout layers. Review spelling, text placement, subject preservation and changes to the reference photo before use. Only the image is returned; requesting IMAGE+TEXT does not add a text output. This tool returns the cover directly without running the separate packaging-presentation stage.
Swap product into a reference ad
Generate an advertisement featuring your product, using an existing ad as the composition and style reference. For a first run, upload a reference advertisement and a clear image of your replacement product. Replace the example Product and copy brief entirely with your product description, campaign intent, desired colors, headline, and supporting copy. Keep aspect ratio on auto and the defaults of 4K resolution and gemini-3-pro-image-preview, then run and inspect the returned image. The reference ad is supplied first for composition, lighting, graphic treatment, and text layout. The product image is resized and padded to the reference dimensions and supplied second. The editing template asks the model to replace the featured product and harmonize lighting and texture. It requests replacement of existing advertising copy while retaining its styling; if the reference has no advertising copy, it requests no added text. Text physically on the product is treated separately from advertising copy. Model selects gemini-3-pro-image-preview or Nano Banana 2 (Gemini 3.1 Flash Image). Resolution offers 1K, 2K, and 4K; the source describes 2K and 4K as using the provider's native upscaler. Auto follows the reference proportions. Explicit aspect ratios request a different output shape, but the editing template also requests the reference ratio, so these instructions can conflict. Under Product reference preparation, choose white or black padding and the product resize method. These affect the product reference supplied to the model; padding color does not directly set the final ad background. Keep the default area resize method for a first run. Uploaded alpha masks are not connected, so these inputs do not preserve transparency or supply a replacement mask. Seed -1 requests a new random seed each run. A fixed seed requests best-effort consistency and does not guarantee identical results. Provider response format offers IMAGE or IMAGE+TEXT; only the generated image is delivered by this Pipe, even when accompanying provider text is requested. This is a generated reinterpretation rather than a precise masked composite. Layout, lettering, branding, requested wording, and product fidelity can vary. Review the result before using it.
Qwen Image Edit
Create an edited version of an uploaded image with Qwen Image Edit. Upload your source image, then describe what should change and what should stay the same. Replace the supplied underwater-scene example with instructions for your own image. For a first run, leave Lightning mode off and use the defaults: 20 standard steps, CFG 2.5, Euler sampler, simple scheduler and Denoise 1. Lightning mode selects the fixed four-step Lightning LoRA together with its separate steps and CFG controls, initially 4 and 1. Only the steps and CFG for the selected mode affect the result. Lightning LoRA strength affects only Lightning mode; guidance normalization and sampling shift apply in both modes. Their supplied defaults are 1 for LoRA strength, 1 for normalization strength, normalization before CFG off, and sampling shift 3. Denoise ranges from 0 to 1: lower values retain more source structure, while 1 allows stronger regeneration. Negative prompt describes unwanted content and may remain empty. Seed -1 chooses a new random seed each run; use a fixed seed to compare settings with the same initial noise. Advanced controls expose the sampler, scheduler and separate sampling settings for each mode. Step counts must be positive integers; the downstream sampler supports 1–10000 steps and CFG values from 0 to 100. The source image supplies both visual conditioning and the starting latent. There is no connected output-size control or automatic 1.5-megapixel resize; resize uploads beforehand if you need a different starting size. The uploaded alpha mask is unused, so source transparency is not preserved. There is no mask control to restrict the edit to a selected area. Editing is generative and may change details beyond the requested changes.
Dieline to box mockup
Create a generated packaging mockup from a flat dieline image and a rendering prompt. For a first run, upload a dieline with its artwork and panel layout visible, leave the Dieline image stage enabled, and run the supplied prompt. It requests an upright box in a three-quarter view on a neutral studio background. Edit Rendering prompt to change the packaging form, camera angle, lighting, or setting. The defaults use Gemini 3 Pro Image, automatic aspect ratio, and 4K target resolution. Auto follows the reference image's proportions; choose an explicit ratio when the flat dieline's shape is unsuitable for the finished composition. The 2K and 4K options use Gemini's native upscaler. Model selects between Gemini 3 Pro Image and Nano Banana 2; their interpretations can differ. Seed -1 requests a new random seed each run. A fixed seed requests best-effort consistency, without guaranteeing identical results. System instructions set general model behavior; the supplied instructions request image generation. For text-only packaging concepts, disable the Dieline image stage and rewrite the prompt to describe the desired packaging and artwork without referring to an uploaded dieline. No upload is needed in this mode. Choose an explicit aspect ratio for a predictable composition shape. Generation uses the Gemini provider and sends the prompt and enabled reference image to that service through a paid provider call. The selected result is the generated image, returned through the Pipe's preview output; a text response is not included. This produces an image, not a dimensional 3D model or a verified folding simulation. Inspect panel placement, artwork, and typography: exact reproduction is not guaranteed, even though the supplied prompt requests it.
ElevenLabs text to speech
Choose a voice and enter the text to speak. Tune speech speed, stability, similarity, and style to shape the delivery. Leave the language code empty for automatic detection. The final output is an MP3; synthesis encoding and final MP3 quality are separate settings. Seed-based reproducibility is not guaranteed.
Photo to geometry maps
Upload one perspective photo to produce two geometry-map images from a shared MoGe-2 estimate. For a first run, upload your image, keep the defaults, and run the Pipe. The first image shows OpenGL normals and the second shows DirectX normals, with inference detail set to 9, automatic camera field of view, and invalid-region filtering enabled. Choose each map independently: OpenGL normals, DirectX normals, depth, colored depth, or the predicted validity mask. OpenGL normals use green = +Y; DirectX normals use green = −Y. Both outputs use the same geometry estimate and inference settings. In Geometry estimation, inference detail ranges from 0 (fastest) to 9 (most detail), as described by the source. Enter the camera's horizontal field of view in degrees when known, or leave 0 to recover it automatically from predicted points. This controls the focal length used to unproject depth into 3D. Invalid-region filtering marks predicted sky and invalid pixels as invalid in points and depth; disable it to retain raw predictions. The predicted mask remains selectable either way. Uploaded transparency is not used as a mask. Automatic resizing checks image width: images wider than 2048 pixels are resized to a 2048-pixel longer edge. Images whose width is at most 2048 remain unchanged, including tall images. Disable automatic resizing to use the original image size. Inference batch size specifies images per inference call, rather than output variations; this Pipe supplies one photo. Force projection is enabled by default, but the supplied source does not explain its effect on the maps. These images are estimates from visible photo content, not guaranteed accurate measurements. This Pipe returns image maps and does not export a mesh or GLB. It requires the fixed moge_2_vitl_normal_fp16.safetensors checkpoint under models/geometry_estimation/.
Flux Kontext image editor
Create one edited image with Flux.1 Kontext Pro using an uploaded image and written instructions. Upload your source image, keep Second image disabled, and try the supplied pop art instruction with the default square output. Describe the subject or region to change and anything you want preserved. The Pipe returns one generated image. To combine two images as editing context, enable Second image and upload another image. Explain in your instruction how each image should be used. Reference layout places the second image relative to the source, matches their adjoining dimensions, and sets the gap and its color. These settings apply only when Second image is enabled. Both images are stitched into a single model input; they are not separate model reference sockets. No second upload is needed while Second image is disabled. Output aspect ratio sets the generated image shape independently of the reference layout. Enter a width:height ratio between 1:4 and 4:1. Guidance controls prompt guidance strength, and Steps sets the number of generation steps; the supplied defaults are 3 and 50. Seed -1 selects a new random seed each run; enter a fixed seed to compare settings. Prompt expansion allows the provider to modify the instruction for more creative generation and makes results nondeterministic even with a fixed seed. This workflow uses a partner API and requires configured access and sufficient credits. Uploaded transparency is not preserved: only the image loaders' RGB image outputs are connected, not their alpha masks.
Seedance 2.0 reference video generator
Generate one Seedance 2.0 video from a prompt, three reference images, a reference video, and reference audio. For a first run, upload all five media inputs and enter your own video prompt. Describe the scene, subject actions, camera movement, and how the references should influence the result. Keep duration at 5 seconds, aspect ratio at portrait 3:4, resolution at 720p, and Generate audio enabled, then run the Pipe. The three images are registered and combined into one reference asset set. They have no assigned start-frame or end-frame roles. Video and audio enter generation as separate references. This connected graph does not establish a mode for omitting references. Image-loader alpha masks are not connected, so uploaded transparency is not preserved as a separate reference mask. Duration is entered in seconds; aspect ratio, resolution, and duration are text fields without an enumerated list of accepted provider values. The supplied defaults are the starting settings. Generate audio changes the generation request and leaves the audio reference connected even when disabled. Seed -1 randomizes each run; a fixed seed does not guarantee identical provider output. Person handling mode and Reference conversion retain source provider inputs whose accepted modes or effects are not documented in the supplied evidence. Keep their defaults for the first run. The result is returned as one video through Sweet Tea. Advanced File type and Codec controls select output encoding; leave both on auto initially. The listed options do not establish support for every container and codec combination.
LTX 2.3 talking and singing portraits
Create talking or singing portrait videos from uploaded audio and portrait images with LTX 2.3. The workflow generates successive segments, spatially upscales and refines them, blends their boundaries, and saves individual segment videos plus an assembled H.264 MP4 at 25 FPS. Both saved outputs use the selected original recording, including accompaniment. For a first run, upload speech or singing and supply an image in all four portrait slots. Reuse the primary portrait in unused slots because every image loader remains connected. Replace shot 1's example prompt, "hello", with a visual description of the subject, setting, expression and motion. Prompts describe visuals; the recording supplies the voice. The source's intended single-shot setup leaves prompts 2–4 empty. Filling them supplies additional image/prompt pairs to the shot looper, but the supplied evidence does not establish its exact empty-slot filtering or rotation order. Audio start time is measured in seconds and defaults to 30; set it to 0 to begin at the start. Duration 0 loads the remaining recording. Select a short duration for an initial trial. Chunk duration is a positive target in whole seconds, not total video duration. One chunker determines loop count from the original recording; another derives segment timings from separated vocals. Both receive the same target duration, but identical boundaries or counts are not established. Their independent variation controls retain the source value of 10; units and exact effects are unspecified. Width and height control shared crop framing for shots 1 and 2, with separate crop positions. Base generation dimensions come from shot 1's crop at half scale, followed by a two-times spatial upscaler; rounding can affect final dimensions. Shots 3 and 4 skip that initial crop, so prepare compatible framing. Only RGB image outputs are connected, so uploaded transparency is not preserved. Generation and refinement have independent image strengths and disable switches. Disabling one stage's image guidance leaves the other stage unchanged and does not remove any upload requirement. Portrait compression is applied before both conditioning stages. The static-camera LoRA affects both sampling stages; strength 0 removes its contribution. Three additional slots accept names of compatible installed LoRAs, each with its own strength. None leaves a slot inactive. Generation steps, schedule shifts, stretching and terminal sigma control the initial sampling schedule. Refinement uses its separate comma-separated sigma sequence. Each stage has an independent seed, CFG value and sampler. Seed -1 randomizes that stage on each run; fix both seeds when comparing other settings. The negative prompt is shared and its influence depends on each stage's CFG. Keep the source sampling defaults for the first run. Decode tiling, overlap and last-frame handling are separate advanced controls; their exact conventions are not fully established by the supplied evidence. Dissolve frames controls boundary blending: 20 frames equals 0.8 seconds at 25 FPS. Overlap trimming is calculated separately, so retain the transition defaults initially. The dissolve curve is a string field; use only values supported by the installed transition node. Final and segment CRF settings independently control H.264 compression. Lower CRF means less compression and larger files. Encoded segments are reloaded for assembly, so final encoding cannot recover detail lost during segment encoding. Final encoding trims the video to the selected audio. Segment outputs are produced within the generation loop and cannot run independently. Longer recordings require more total processing and storage. The workflow relies on its configured models and custom nodes; the supplied dependency catalog does not resolve every required asset.
Wan 2.2 controlled video
Generate a prompted video with Wan 2.2 Fun Control 5B, guided by a source video's frames or extracted edges and an optional reference image. For a first run, upload a source video, replace the example prompt with your desired scene and motion, and either upload a reference image or disable the Reference image stage. That stage is enabled initially, so leaving its upload empty requires disabling it. Keep 704 × 704, 121 frames, 24 FPS, batch size 1, and the default sampling settings. Canny edge guidance is disabled initially: original source frames supply guidance. Enable it to use extracted edges instead. Its normalized low and high thresholds apply only in that mode; keep the low threshold below the high threshold. Reference image guidance can be disabled independently. Generation uses these inputs as conditioning rather than directly editing the source video's pixels. Width and height are pixels in multiples of 16. Frame count uses 4n + 1 values, such as 81 or 121. Playback FPS is independent of generated frame count: 121 frames at 24 FPS yield approximately 5.04 seconds. Changing FPS changes playback duration without requesting more frames. Source audio and source FPS are not carried into the output. Advanced controls include unwanted traits in the negative prompt, seed, sampling steps, prompt guidance, sampler, scheduler, denoise fraction, and sampling shift. Seed -1 randomizes each run; a fixed nonnegative seed reuses sampling noise. Denoise is not a direct source-video preservation strength. The retained negative prompt is the source's Chinese-language list of unwanted traits and can be edited. Video playback offers 8-bit or 10-bit creation; 10-bit support depends on downstream encoding and players. Video output offers container and codec choices; leave both on auto initially. The graph returns one video output with no separate edge preview. Batch size changes the latent batch but does not provide separate output branches for individual clips, so keep it at 1 for an ordinary single-clip run.
Gemini image edit and generation
Edit a reference image or generate images from text using Gemini. For a first run, upload a source image, keep the Source image stage enabled, and describe the changes you want and the features to retain. The supplied example requests anime styling while retaining identity. Keep the defaults: Gemini 3 Pro Image Preview, Auto aspect ratio, 1K resolution and seed -1 for randomization. For text-only generation, disable the Source image stage and replace the editing instruction with a complete scene description. An upload is required while the source stage is enabled. Auto aspect ratio follows the reference when supplied; choose an explicit ratio for text-only generation when output shape matters. Resolution selects a 1K, 2K or 4K target. The source uses Gemini's native upscaler for 2K and 4K. Model selects between Gemini 3 Pro Image Preview and Nano Banana 2 (Gemini 3.1 Flash Image). A fixed nonnegative seed offers best-effort repeatability, but identical results are not guaranteed, including when model or request settings change. System instructions set foundational behavior alongside your image instruction. The supplied instructions direct the model to produce images and interpret requests visually. Model response mode selects IMAGE or IMAGE+TEXT; this Pipe returns images only, so any accompanying text is not surfaced. This tool connects one reference image and has no mask-based editing input or multi-reference batching stage. Uploaded alpha transparency is not preserved because the loader's separate alpha-mask output is not connected.
