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
17 affiché of 41 736
SeedVR2 image upscale
Upscale and restore an uploaded still image with SeedVR2, returning the restored image directly. For a first run, upload an image, leave the target width at 1080 pixels and guidance at 1.01, then run. Compare the result with your original: restoration can reinterpret fine details. The target-width control accepts 1–4320 pixels. The seed defaults to -1 for a fresh random seed each run. Choose a fixed seed from 0 to 5000 when comparing guidance or other settings. Guidance accepts 0.01–2. GPU-memory conservation is enabled by default and can be toggled. Processing batch size defaults to 1; keep that starting value for the single uploaded image. It does not select the number of output variants. The fixed seedvr2_ema_3b_fp16.safetensors model and the SeedVR2 custom node must be available in the runtime. This Pipe accepts still images and has no video input or encoding mode. Uploaded transparency is not preserved because the loader's separate alpha-mask output is not connected.
veryBadImageNegative
veryBadImageNegativev1.3 uses a new training atlas and performs well when used in AOM3 and viewer-mixv1.7 .
LongCat Image Edit
Create an edited image with LongCat Image Edit by providing a source image and describing the desired change. Set output resolution in megapixels while retaining the source aspect ratio. Advanced controls include edit strength, positive and negative guidance, sampling settings, and seed. An optional negative prompt describes elements to avoid.
Vidu Q3 text to video
Generate one short video from a text description of the scene, action, camera movement and visual style, with optional generated sound. No uploaded media is required. This uses the Vidu partner API, requires authenticated provider access and makes a paid provider call; your prompt is sent to the provider. For a first run, use the included neon street-chase prompt with Pro, six seconds, 16:9 framing, 720p resolution and sound enabled. Replace the prompt to create your own scene. Keep it within 2000 characters and align any requested shot timings with the selected duration. Choose Pro or Turbo, one of five aspect ratios, 720p or 1080p resolution, and a duration of 1–16 whole seconds. Enable sound when describing desired dialogue or sound effects; disable it for silent video. Prompt details express the desired result and do not guarantee exact shot timing or dialogue. Leave seed at -1 for a fresh random seed each run, or enter a fixed seed. The generated video is returned for preview with automatic format and codec selection.
JoyAI Image Edit
Generate edited versions of a reference image using natural-language instructions. For a first run, upload the image you want to edit, replace the example instruction with your requested change, and state what should remain unchanged. For example: “Change the background to a glacial scene; keep the bottle and its label.” Leave Output resolution at 1 megapixel, Variations at 1, and the sampling defaults unchanged, then run. Output resolution sets the approximate total pixel count. The reference is resized before conditioning, and its resized dimensions determine the generated image dimensions, aligned to multiples of 16. Resize method changes how the reference is resampled. Increasing resolution or the number of variations increases memory use. This is a generated edit, not a masked, pixel-preserving retouch: changes can extend beyond the requested region. Uploaded transparency is not preserved because only the loader's RGB IMAGE output is connected. Negative prompt describes unwanted content and can remain empty. Both positive and negative conditioning receive the resized reference. Variations generates multiple edited images in one batch. Seed -1 selects a new random seed each run; use a fixed seed when comparing control changes. Advanced sampling controls include Steps, CFG, Sampler and Scheduler. Start with 40 steps, CFG 4, Euler and Normal. Denoise controls the sampling schedule fraction; retain 1 initially. Since generation starts from an empty latent, lowering denoise is not a direct reference-preservation control. Guidance normalization exposes independent strength and timing settings; begin with strength 1 and Normalize before CFG enabled. The Pipe returns the edited image batch through Sweet Tea's image output. It does not return a separate side-by-side comparison. The original save node has been replaced, so its file-format and permanent-save settings do not describe this selected output.
GPT Image generation and editing
Create one or more images using a text prompt, optionally guided by an uploaded reference image. For a first run, leave Reference image disabled, enter your prompt, and use the default square size, high quality, and one image. To edit an image, enable Reference image, upload it, and describe the changes in the prompt. Output size selects a preset or automatic sizing. Custom width and height apply only when Output size is Custom and Model is gpt-image-2; each dimension must be 1024–3840 pixels in multiples of 16. Quality affects generation time and cost, and Image count requests 1–8 results. Model offers the three source-supported GPT Image variants. Background requests automatic, opaque, or transparent output from the provider; the supplied graph does not establish preservation of output alpha. Uploaded reference transparency is discarded by the RGB input path, and this tool has no mask-based inpainting interface. Seed is not implemented by the provider backend and cannot reproduce a result. Runs require access to the ComfyUI partner API with authentication and sufficient credits.
ZIB + ZIT image generator and upscaler
Generate a final image batch from text using the fixed Moody ZIB composition model and Moody ZIT detail model, followed by tiled enlargement. No upload is required. For a first run, replace Positive prompt with your subject, composition, lighting and style. Keep the 640 × 960 starting canvas, Image count of 1 and supplied sampling settings. Leave Face detailing and Contrast finishing disabled. The example prompts favor cool, pale portrait styling; replace the tile and face examples when enabling those options for another subject. Detail enlargement defaults to 1.7× before the second diffusion pass. Final enlargement defaults to 2× during tiled refinement and supports 0.05–4 in increments of 0.05. Final dimensions are approximately the starting dimensions multiplied by both factors, subject to latent rounding. Factors below 1 reduce dimensions. Larger canvases, factors and image counts increase memory use and processing time. This Pipe always generates its starting image and has no uploaded-image input for standalone upscaling. Positive prompt guides both generation passes and normally tile refinement. Enable Use separate tile prompt to substitute Tile prompt for the tiled stage. Negative prompt affects only composition; later diffusion stages use zeroed negative conditioning. Tile denoise controls how strongly the enlarged image is repainted. Composition steps defines a full 17-step schedule, with Composition start step 0 and Composition stop step 12. The partial pass retains noise before latent enlargement. Detail steps defines a separate 11-step schedule starting at Detail start step 5; Detail stop step 999 allows it to finish. Start and stop indices belong to each stage's own schedule. Keep schedule lengths, boundaries and noise settings consistent when changing them. Latent resize method changes interpolation between these passes. Composition sampling shift affects the first model; Detail and tile sampling shift affects both subsequent stages. Composition, detail, tile and optional face sampling have independent seeds. -1 randomizes each active stage on each run. Fix all active stage seeds to reuse their randomization; other control changes can still change the result. Tile dimensions are calculated automatically from the detail-pass image dimensions and Final enlargement. The formulas retain a fixed 96-pixel term even if Tile padding changes, so preserve the padding default initially. Tiles per batch controls tile processing rather than output image count, and applies to redraw rather than seam correction. Seam correction starts at None; seam denoise, band width, blur and padding apply when a correction mode is selected. Tiled decoding controls VAE decoding separately from image redraw tiling. Enable Face detailing to detect and repaint faces after enlargement. Face prompt is independent of Positive prompt. Face denoise controls how strongly faces change and may affect identity or expression. Face crop sizing, detection, segmentation, noise masks, sampling and encoding/decoding controls apply only while this stage is enabled. Guide size works with its bounding-box or crop basis and maximum size. Keep Face inpainting model mode false for the fixed supplied model, and leave Face wildcard text empty unless using supported Impact Pack syntax. Enable Contrast finishing to apply the fixed 1× skin-contrast model, then adjust Contrast strength and Blend mode. With normal blending, strength 0 keeps the pre-finish image and 1 uses the finish. When finishing is disabled, normal blending leaves the image unchanged regardless of strength. Other blend modes can alter the image even while finishing is disabled. The model pair and source LoRA configurations are fixed. Manual masked detailing is unavailable because its mask depends on a live ComfyUI editor session rather than a portable Pipe input. The retained manual bridge and regional detailer remain bypassed without enable toggles; their archived settings do not affect the output. Only final images are returned.
Wan 2.1 camera-controlled image to video
Animate an uploaded still image into a prompt-guided video with selectable camera movement using Wan 2.1 Fun Camera 1.3B. For a first run, upload a source image, replace the landscape example prompt with a description of your scene and desired subject action, and choose a camera movement. Keep the remaining defaults initially: Zoom In, camera speed 1, 512 × 512 pixels, 81 frames, 16 FPS and generation batch size 1. These settings produce about 5.1 seconds of silent video. The source image supplies starting-frame conditioning and a separate visual reference; its alpha mask is not connected, so transparency is not preserved. Camera choices include Static, directional pans, zooms and clockwise or counterclockwise rotation. Camera speed changes camera conditioning. Width and height use multiples of 16; frame count follows 1, 5, 9 and so on. These shape controls drive both camera and video conditioning. Playback FPS changes duration without changing the generated frame count. Static camera conditioning still permits subject motion described by the prompt. Camera controls condition generation rather than guaranteeing an exact rendered trajectory. In advanced settings, edit the negative prompt to describe unwanted results. Image guidance can be bypassed while keeping the uploaded image connected as the starting frame. Reference crop applies only to image guidance and has no effect when that stage is bypassed. Camera focal-length and optical-center controls are available in Camera and video shape; their source uses a 0–1 scale. Sampling controls include seed, steps, prompt guidance, sampler, scheduler, denoising amount and sampling shift. Seed -1 selects random noise; use a fixed nonnegative seed to reuse the initial noise with the same settings. Denoising amount is not a direct image-preservation slider. The saved sampling settings are 20 steps, CFG 6, uni_pc, simple scheduler, denoising amount 1 and shift 8. Video playback and export expose FPS, 8- or 10-bit depth, container, codec and CRF. Start with automatic container and codec and CRF -1. Some players or downstream tools may not support 10-bit video. Generation batch size controls latent batching, not separate exported files; the selected output assembles decoded images into one video, so use 1 for a single clip. This tool has no connected audio input and returns the video output rather than an animated WebP or individual frame export.
Grok video extension
Continue an existing video with a prompted 2–10 second extension using Grok. For a first run, upload an MP4 clip lasting 2–15 seconds, replace the example space-scene prompt with what should happen next in your footage, and keep the default six-second extension. Describe the action, setting, and camera movement that should follow the clip’s ending. Extension length requests 2–10 seconds in whole-second increments; it does not set the source clip’s length or total output length. The Pipe returns the video supplied by Grok, with no separate concatenation stage. Advanced output encoding controls offer file type, codec, and CRF settings; keep their defaults for a first run. The Grok partner API requires runtime authentication. Leave Seed at -1 for randomization on each run. Seed is a rerun control, and a fixed value does not make Grok’s results reproducible.
Whole-body pose map
Create a rendered whole-body pose map from an image of one person using SDPose. Upload your source image, leave the longer edge at 1024 pixels and all drawing regions enabled, then run to view the output image. The input is resized with its aspect ratio preserved before pose extraction. Choose whether to draw the body, hands, face, feet and head. Head drawing is separate from facial keypoints. Line width changes the skeleton's thickness; face point size applies when Draw face is enabled. The confidence threshold ranges from 0 to 1: raising it omits less-confident keypoints, while lowering it can display more uncertain detections. These controls affect the rendered map rather than pose extraction. Longer edge sets the resized image's longer dimension in pixels; use a positive value. Resampling method selects interpolation before detection. Area is suited to downscaling and Lanczos to upscaling. Advanced settings expose the extractor's processing batch size, defaulting to 16; this does not request multiple output variations. This tool uses the fixed sdpose_wholebody_fp16.safetensors checkpoint and returns a pose visualization through the image preview output, without a separate keypoint-data output. The extractor requires bounding boxes for multi-person detection, and this workflow supplies none, so use a single-person image.
Gemini Omni Flash text to video
Generate a 3–10 second video with audio using Google's Gemini Omni Flash model. The source node specifies 720p output at 24 FPS. For a first run, keep the example video prompt, portrait aspect ratio (9:16), three-second duration and default advanced settings, then run the Pipe. No image, video or audio upload is required. To create your own clip, describe the scene, action, visual style, camera movement and desired sound. Use 9:16 for portrait or 16:9 for landscape. Enter duration as 3s through 10s in whole-second increments. These controls append instructions to the scene prompt rather than cropping or trimming the result; avoid conflicting duration or framing instructions in your prompt. Advanced generation settings expose temperature and top-p sampling controls. Seed -1 randomizes the request seed each run. Seed affects whether the generation node reruns, but results remain non-deterministic even when a seed is repeated. Advanced video output settings expose file type, codec and CRF for output transport. Keep their defaults for the first run; these settings do not change the generation model or scene instructions. Generation requires access to the ComfyUI partner API. The workflow sends the assembled prompt to the Gemini provider and makes a provider-priced generation call. This Pipe uses text input only and returns the generated video; the model's separate text output is not selected.
Wan image to video
Animate an uploaded image into one video using Wan 2.5 or Wan 2.6, with optional generated audio. Upload the image you want to animate and replace the example prompt with instructions matching its subjects and setting. Describe the scene, subject motion, and camera movement in English or Chinese. For a first run, keep Wan 2.5, 720P, and 5 seconds. The result appears in Sweet Tea's video output. Choose 480P, 720P, or 1080P and a duration of 5, 10, or 15 seconds. A 15-second video requires Wan 2.6. The negative prompt describes elements to avoid and may be left blank. AI prompt extension is enabled by default and lets the provider enhance your instructions; disable it to submit your wording without that enhancement. Generated audio is enabled by default and can be turned off. This Pipe has no uploaded audio input. Watermarking is optional and disabled by default. Use seed -1 to randomize each run, or enter a fixed generation seed; identical provider results are not guaranteed. Advanced video output controls select the file type and codec. Leave both on auto for the first run and choose compatible settings when changing them. Uploaded transparency is not preserved: only the loader's RGB image output feeds generation. Generation sends the image and prompt to the Wan API and uses a provider-priced generation call.
Video frame interpolation
Add intermediate frames to an uploaded video for smoother playback or slow motion, returning one video at the source frame dimensions. For a first run, upload a source video, keep Frame multiplier at 2 and Keep original speed enabled, and use the default film_net_fp16.safetensors interpolation model. Leave the advanced bit depth and encoding settings at their defaults. With Keep original speed enabled, output FPS is the source FPS multiplied by Frame multiplier, giving smoother playback at approximately the original duration. Disable it to retain the source FPS and play the additional frames as slow motion. The multiplier supports integers from 2 to 16; larger values create more frames and increase processing work. Output FPS is calculated automatically, and there is no resizing control. The advanced Interpolation model control selects from the source loader's available runtime models. Models must be available in the runtime's frame_interpolation folder; the default is film_net_fp16.safetensors. Source audio is passed through unchanged. It is not slowed or stretched, so slow-motion output may have mismatched audio and video durations. Advanced controls select 8-bit or 10-bit video, the output container (Auto, MP4, MKV or WebM), and codec (Auto, H.264 or AV1). Start with 8-bit and automatic format and codec selection. The source describes 10-bit as reducing gradient banding, but some players and downstream tools may not support it. Explicit container and codec choices must be supported together by the runtime and your playback software.
Wan 2.1 Camera Motion
Animate an uploaded image into a silent video using Wan 2.1 Fun Camera 14B, a scene prompt and explicit camera-motion conditioning. For a first run, upload your source image, replace the example jazz-bar prompt with a description of its scene, subject action, lighting and style, and choose a camera movement. Keep camera speed and sampling settings at their defaults. The default output is 512 × 512 pixels with 81 frames at 16 FPS, approximately 5.1 seconds. Use dimensions in multiples of 16 and frame counts of 4n + 1, such as 49, 81 or 121. Available movements are Static, directional pans, zoom in or out, and clockwise or counterclockwise rotation. Camera speed is a relative conditioning parameter, not a physical speed. Align camera instructions in your prompt with the chosen movement. Static requests a stationary camera; subjects can still move. Generated motion is model-guided and does not guarantee a geometric reconstruction of the scene. Frames determines how many frames are generated. FPS sets playback rate: raising FPS shortens playback without adding frames. Width, height and frame count are supplied by Camera and output shape. The uploaded image supplies both the starting image and visual-reference conditioning and is required in this graph. Reference crop changes only visual-reference encoding. Uploaded transparency is not retained through the RGB image path. Advanced controls include negative prompt, seed, sampling steps, prompt guidance, sampler, scheduler, sampling shift and denoise. Use seed -1 for a new random seed each run, or a fixed seed when comparing changes. Denoise is a sampling-schedule fraction rather than a direct image-preservation strength. Optional normalized focal and principal-point parameters adjust camera intrinsics; retain their defaults initially. Keep batch size at 1 for one clip: this Pipe has one assembled video output rather than separate outputs for each latent sample. Video playback and encoding controls provide FPS, 8-bit or 10-bit video, container, codec and CRF. Retain automatic container and codec selection for the first run. Ten-bit output requires compatible encoding and playback support. CRF changes encoding quality rather than generated content. No audio is connected, so the output is silent.
Kling reference image to video
Generate one video from a reference image and a motion prompt. Upload one image, then replace the example prompt with the desired action, scene and camera movement. For a first run, keep kling-v3-omni, five seconds, 16:9 framing and 720p resolution. Generated audio is enabled by default. Choose landscape (16:9), portrait (9:16) or square (1:1) framing, a duration from 3 to 15 seconds, and 720p, 1080p or 4k resolution. The Model control also offers kling-video-o1; turn Generate audio off when selecting it because generated audio is supported only by kling-v3-omni. Execution uses the ComfyUI partner API, sends the reference image and prompt to Kling, and incurs provider charges. Model, duration, resolution and audio choices affect the charge. The image guides the subject and appearance; this Pipe has no separate first-frame or end-frame input. Include desired content and things to avoid in the same prompt. Storyboard mode remains disabled because this interface has no editable segment prompts or durations. Seed defaults to -1 for randomization. It controls provider re-execution rather than reproducibility: results remain non-deterministic with a fixed seed. The Pipe returns one video for preview. Advanced Video output settings provide file type and codec choices, both defaulting to automatic selection.
Flux 1.1 Pro Ultra image generator
Create a single image with Flux 1.1 Pro Ultra. For a first run, edit the prompt, choose an aspect ratio, and leave reference image guidance disabled. The default produces a square image without an uploaded input. To guide generation with an existing image, enable Reference image and upload an image. Reference strength controls its influence: higher values make the result closer to the reference. This setting has no effect while Reference image is disabled. Enable Raw mode for less processed, more natural-looking images. Prompt expansion lets the provider modify the prompt for more creative generation; with it enabled, the same seed will not produce exactly the same result. Seed defaults to a new random value each run. This workflow uses a hosted API node and requires configured API access and sufficient credits. Output size is controlled by aspect ratio; the source does not expose explicit pixel dimensions.
Boogu image edit
Create edited images from one uploaded reference and a natural language instruction. Upload your source image, enter the change you want and run with the defaults: 1 megapixel, one variant, 25 steps and instruction guidance (CFG) 3.5. “Remove the hat” is an example instruction; choose wording appropriate to your image. The negative prompt can describe unwanted details, or remain empty. Output size sets a target pixel budget in millions while retaining the source aspect ratio. Bypass Output size to derive dimensions from the source image instead. This stage controls output dimensions; the original upload still feeds editing conditioning. Resampling method affects only the image used to derive dimensions. Variants sets the number of edited images returned in a batch. Larger dimensions and batches require more memory. Advanced controls include guidance, seed, sampling steps, sampler and scheduler. Seed -1 selects fresh randomness each run; use a fixed seed when comparing settings with Add noise enabled. Leave Add noise enabled and Noise schedule fraction at 1 for a first run. The workflow starts from an empty latent conditioned on your image and instruction, so lowering the schedule fraction does not directly preserve source pixels. Schedule adjustment changes the sampling configuration used to calculate noise levels; the sampler itself receives the original model. This is a generative editor without a connected editing mask. Instructions to retain details do not guarantee that all other content stays unchanged. Uploaded transparency is not retained because only LoadImage’s IMAGE output is connected. The fixed model files are boogu_image_edit_fp8_scaled.safetensors, qwen3vl_8b_fp8_scaled.safetensors and ae.safetensors. Supplied dependency evidence marks all three model files and the Sweet Tea output node pack as missing in the inspected environment; their source locations are supplied. The selected result is the decoded edited-image batch returned through Sweet Tea’s image output.
