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
19 affiché of 42 416
Kling two-shot storyboard
Generate one Kling video containing two consecutive storyboard shots, with optional first-frame guidance and native audio. For a first run, replace both shot prompts with your subjects, action, setting and camera movement. Set each shot's duration in seconds; the defaults request two seconds followed by three seconds. Each prompt supports up to 512 characters, and each shot duration supports 1–15 seconds. These controls request the segments and their combined duration; the example prompts describe a boxing sequence. Upload a start image and leave Start image enabled for image-to-video generation. An upload is required while that stage is enabled. For text-to-video generation, bypass Start image; no upload is required. Aspect ratio offers landscape 16:9, portrait 9:16 or square 1:1 in text mode and is ignored when a start image is supplied. Kling V3 is selected by default and supports 720p, 1080p and 4k, with optional native audio. You can also choose Kling 3.0 Turbo, but select 720p or 1080p before running it. Turbo always generates native audio and ignores the audio toggle. This tool supplies exactly two storyboard segments; other segment counts and ordinary single-prompt generation are not configured here. Leave Seed at -1 for a fresh run. A fixed seed does not make the provider's output reproducible. Generation uses authenticated Kling API access, sends prompts and any supplied start image to the provider, and incurs provider charges. The result is returned as one video. Advanced Video output controls offer file type and codec selection; leave both at auto for automatic selection, or choose a compatible explicit combination.
Corneo's Marin Kitagawa TI Embedding
This is a Textual Inversion Embedding of Marin Kitagawa from My Dress-Up Darling (sono bisque doll wa koi wo suru).
Audio-guided human video
Generate one human-focused video from source audio and positive and negative prompts, optionally guided by a reference image. The uploaded audio supplies HuMo conditioning and is also included as the soundtrack. For a first run, upload the desired audio excerpt and describe the person, clothing, setting, framing, and action in Positive prompt. Retain or edit Negative prompt to specify unwanted visual features; its supplied default is in Chinese. Upload a reference image to guide appearance and composition, or bypass Reference image to use audio and prompts alone. The reference stage starts enabled, so explicitly bypass it when no image is supplied. Start with 640 × 640 pixels, 97 frames, 25 FPS, and latent batch size 1. The generated frames represent approximately 3.88 seconds of playback. Prepare an audio excerpt appropriate for that span before uploading: this interface provides no verified audio trimming or time-stretching controls. Width and height use multiples of 16. Frame count follows 1 plus a multiple of 4, such as 97, 101, or 105. Frame count controls the number of generated frames; frame rate controls their playback duration. Changing frame rate does not change audio encoding or stretch the soundtrack. Optional 10-bit video can retain smoother gradients, but downstream codec and player support varies; the default is 8-bit. Advanced sampling exposes seed, steps, CFG, sampler, scheduler, denoise fraction, and sampling shift. Seed -1 randomizes each run; a fixed seed reuses sampling noise. The supplied preset uses six steps, CFG 1, uni_pc, simple scheduling, denoise 1, and shift 8. The fixed Lightx2v step-distillation adapter starts at strength 1; its strength is not a dedicated motion control. Keep these settings initially. The model, text encoder, audio encoder, and VAE remain fixed. The result is one video with the uploaded audio, with automatic file type and codec by default. Advanced Video output settings expose container and codec choices. Larger latent batches still enter the same video assembly rather than producing separate clip outputs. There is no separate image or audio output.
SeedVR2 video restore and upscale
Restore and resize an uploaded video with the fixed SeedVR2 3B Int8 model. The default run processes the full clip at 2× its source width and height with one sampling step. No text prompt is required: conditioning comes from the video itself. The full clip or selected segment supplies the output frame rate, audio when present, and bit depth to video assembly; final encoding also depends on the output settings. For a first run, upload a short video and keep the defaults. Enable Trim video to process a segment. Start time and Duration are in seconds; Duration 0 continues to the end. Require full duration raises an error if the requested duration cannot be provided. Segment settings apply only when trimming is enabled. Scale factor changes both dimensions while preserving aspect ratio: 2 doubles width and height, 1 retains the source dimensions, and values below 1 reduce them. A factor of 2 produces approximately four times as many pixels per frame. Resampling method controls interpolation before restoration; the source recommends lanczos for upscaling, area for downscaling, and nearest-exact for pixel art. Color matching optionally brings restored colors closer to the resized source. lab transfers color in CIELAB space, wavelet transfers low-frequency color, adain matches global channel statistics, and none skips color transfer. Enable Temporal chunking to sample automatically sized latent chunks, then merge them before decoding. Chunk overlap is measured in latent frames, rather than seconds or source video frames; overlapping chunks are crossfaded during merging, while 0 uses plain concatenation. It applies only when chunking is enabled. Chunk sizing is automatic; manual chunk size is not available in this Pipe. Chunking addresses sampling memory, while VAE encoding and decoding have their own spatial tile and temporal window controls. Long clips and large output dimensions can still require substantial memory; use a shorter segment when needed. Denoise 1, one sampling step, CFG 1, euler, and simple are the source restoration defaults. Lower Denoise values change the sampling schedule and may reduce changes; test on a short segment. Other sampling settings depart from the source setup and are experimental. Seed -1 selects a new random seed each run; use a fixed seed when comparing settings. Video file type and Codec default to automatic selection. If choosing them explicitly, use a compatible container and codec supported by your intended player. Quality (CRF) -1 keeps automatic selection; lower nonnegative values request higher encoding quality and generally larger files, with the effective range depending on the codec.
Portrait lighting transfer
Generate a relit portrait or subject image with Qwen Image Edit and the Light Migration LoRA. For a first run, upload the required subject image and a lighting reference, keep the default relighting instruction, and leave Lightning mode off. Start with standard steps 20, standard CFG 4, Light Migration strength 1 and denoise 1. Run the Pipe to receive one edited image. Add subject details you want retained to the positive prompt; the negative prompt can remain blank. For text-directed lighting, bypass the Lighting reference stage and replace the default instruction with a direct description of the desired illumination and colors. Leaving the upload empty alone does not select this mode. The subject image remains required. Lightning mode adds the supplied Qwen Image Edit 2509 four-step Lightning LoRA and switches steps and CFG together. Its defaults are four steps and CFG 1. Standard steps and CFG have no effect while Lightning mode is on; Lightning steps, CFG and LoRA strength have no effect while it is off. Light Migration strength applies in both modes. The required model files must be available, including the four-step Lightning LoRA when using that mode. Denoise ranges from 0 to 1 and controls regeneration of the initial subject latent. Lower values retain more source structure but may weaken the lighting change. Light Migration strength controls the lighting LoRA's model influence, rather than an exact percentage of transferred lighting. Use seed -1 for a fresh random seed or a fixed seed to compare edits with otherwise unchanged settings. Advanced controls include separate steps and CFG for each mode, Lightning LoRA strength, sampler, scheduler, guidance normalization and sampling shift. The source sampler and scheduler defaults are Euler and Simple. Steps must be 1–10000 and CFG 0–100, matching the downstream sampler. Guidance normalization and sampling shift apply in both modes; normalization before CFG can amplify guidance. The subject is automatically scaled before generation, so output dimensions may differ from the upload. There is no independent output-size control. This is generative editing: identity, edges and fine details may change. Uploaded transparency is not preserved.
Style Rust Magic
This is a textual inversion that I made inspired by The Last of Us, created with Dream Artist on a single image that I generated. I was trying to do a post-apocalyptic feel. It adds a lot of gadgets, rusted...
Wan 2.2 5B video generation
Generate a silent video with Wan 2.2 5B from a scene and motion description, optionally using an uploaded image to condition the opening. For a first run, leave Source image bypassed and replace the positive prompt with your scene, subject movement, visual style, and camera movement. Keep the supplied negative prompt and default 1280 × 704 dimensions, 121 frames, 24 FPS, and 20 sampling steps. These settings produce approximately five seconds of video. Seed -1 selects fresh random noise each run; use a fixed seed when comparing changes. To animate an image, enable Source image and upload an opening image. Describe the desired movement and continuation in the positive prompt. The upload connects to the opening-image input, rather than supplying a video for editing. Uploaded transparency is not preserved through the connected RGB image path. Width and height are measured in pixels and use multiples of 32. Frame counts use 4n + 1, such as 49, 81, or 121. Increasing dimensions or frame count increases generation work. FPS controls playback speed: with a fixed frame count, increasing FPS shortens the video rather than generating additional frames. Keep batch size at 1 for one sequence; larger batches still feed one video output and do not provide separate video files. The negative prompt discourages unwanted content and artifacts. Advanced Sampling and Noise schedule controls change the denoising process; begin with their supplied defaults and denoise 1. Lower denoise is not a direct opening-image fidelity control. Use 8-bit video for broader playback compatibility. The 10-bit option requires compatible encoding and players. Advanced Video output controls select container, codec, and compression quality. Start with automatic file type and codec and quality -1; explicit encoding choices must be compatible. Encoding quality affects compression, not generated scene content. The output has no audio, even when the prompt describes music or sounds.
SCG Embedding Toolkit for Version 1.5 Models - A collection of tools and helpers from my personal kit
Over the last 6 months, I've created quite the collection of helper and tool embeds for doing things like fixing faces, pushing realism, creating anime scenes, making subjects smile, or cry, or face the camera,...
SoCalGuitarist Cinema768 Embeddings - Digital / Analog / Classic B&W - Easy movie scenes for 2.X Models!
These are another batch of high quality SD model 2.X embeddings that I am excited to release to the stable diffusion community! My new Cinema768 embeddings deliver easy cinematic lighting, color grading and...
bad_pictures
A very versatile negative embedding that helps generate high quality images. Higher chances of producing normal looking hands in realistic images.
Flux 2 Klein four-step text to image
Create images from a text description using the fixed distilled Flux 2 Klein 4B model. For a first run, replace the example Positive prompt with your subject, setting, composition, lighting and desired visual style. No uploaded image is required. Keep the source defaults: 1024 × 1024 pixels, one image, four steps, CFG 1 and Euler sampling. Run the Pipe to receive the generated image in Sweet Tea. Width and height set both the image canvas and the resolution used by the sampling schedule. Enter positive multiples of 16, at least 16 pixels. Together they determine the aspect ratio. Images per run generates a batch with the same prompt and dimensions; larger dimensions and batches require more memory. Seed -1 selects fresh random starting noise each run. Use a fixed nonnegative seed to compare changes with the same starting noise. Advanced controls expose steps, CFG and sampler. Steps set the sampling iteration count; additional steps add computation and do not guarantee better results. Keep four steps and Euler for the source setup. Changing the sampler can change results and interact with the step count. CFG adjusts positive conditioning relative to zeroed conditioning. Keep 1 for the source setup; higher values change that guidance. This branch has no editable negative prompt. This Pipe generates only the distilled-model branch's images. The separate Klein Base model with editable negative prompting belongs to the sibling tool and is not a selectable mode here. The runtime requires flux-2-klein-4b.safetensors in diffusion_models, qwen_3_4b.safetensors in text_encoders, and flux2-vae.safetensors in vae. These model dependencies are fixed. The output transport returns images to Sweet Tea without permanently saving them through ComfyUI.
Kling motion transfer
Animate a character image using movement and expressions from a reference video, returning one generated video. For a first run, upload a character image and a 3–10 second motion reference video. Keep Orientation source set to video, replace the example Scene prompt with your character, setting and desired details, then run the Pipe. With video orientation, movement, expressions, camera motion and character orientation follow the reference video; supported reference duration is 3–30 seconds. With image orientation, movement and expressions still follow the video, character orientation follows the image, and camera details come from your prompt; supported reference duration is 3–10 seconds. Keep reference audio determines whether the motion video's original sound is retained and is enabled by default. Generation mode offers pro and std, while Model version offers kling-v3 and kling-v2-6. Defaults are pro and kling-v3; the supplied source does not establish quality or performance differences between these choices. Advanced Video output settings let you choose a file type and codec; both default to automatic selection. The CRF encoding control defaults to -1, and its codec-dependent behavior is not documented in the supplied evidence. Generation uses the Kling provider and transfers the reference media for a provider-priced request. Uploaded image transparency is not preserved because only the loader's IMAGE output is connected.
Styled logo rotation
Generate a styled logo animation from a logo image and a material or matcap reference. The material reference is resized and padded to the logo's dimensions, then both images guide generation of a styled first frame. Gemini 2.5 Flash Image generates the last frame from that first frame, and Seedance animates between the two generated endpoints. For a first run, upload both required images and keep the default prompts and five-second duration. These request a studio-lit logo, a matching back view, and a 180-degree rotation. Prepare the logo's background before uploading: uploaded transparency is not carried through the connected RGB image inputs. In the styling prompt, image 1 is the logo and image 2 is the material reference. In the last-frame prompt, image 1 is the generated styled first frame. To change the movement, edit the last-frame and motion prompts together so the destination view agrees with the requested motion. Fixed camera adds a stationary-camera instruction; the default motion prompt also explicitly requests a static camera. If you want camera movement, disable Fixed camera and revise that prompt. First-frame model selects gemini-3-pro-image-preview or Nano Banana 2 (Gemini 3.1 Flash Image). First-frame shape and resolution govern the initial image; the last frame defaults to matching its shape. Video shape and resolution govern the final clip independently. Duration is 3–12 seconds. Generate audio applies only to the Seedance 1.5 Pro model. AI watermark requests a watermark from the video service. Reference preparation offers white or black padding and a resampling method. Video export offers container and codec choices; keep their Auto defaults for a first run. Both endpoint images are generated on every run; uploaded endpoint frames are not supported. Each generation stage has its own seed: -1 randomizes that stage. Fixing only the video seed does not fix the generated endpoints. Fixed Gemini seeds request best-effort repeatability rather than deterministic results. Geometry, lettering, material consistency, and camera stability are not guaranteed. The selected result is the video; model text responses are not connected to an output.
Color Palettes
Easy one-click color palette for your images. No need to argue with the AI if an orange is a color or a fruit, everything is setup for you automatically.
FlashVSR video upscale
Upscale an uploaded video to 720p, 1080p, 2K, or 4K with WaveSpeed FlashVSR and return one processed video. For a first run, upload your source video, keep Target resolution at 1080p, and leave File type and Codec on auto. Run the Pipe to receive the upscaled video. The uploaded video is sent to WaveSpeed for processing through a paid provider call. Target resolution controls the FlashVSR request. Under advanced Video output settings, choose auto, MP4, MKV, or WebM for the container and auto, H.264, or AV1 for the codec. These controls affect the returned video's encoding rather than the requested upscale resolution; use automatic encoding unless you need a specific format. This Pipe processes an existing video. It exposes no prompts, restoration strength, frame-rate changes, or duration controls. The source CRF default is retained as a hidden encoding setting because its behavior is not documented in the supplied evidence.
赛博世界/Cyberworld/赛博朋克场景 Lora
推荐的质量词:masterpiece, best quality, score7, score8, score9, highres, extremely detailed 8k wallpaper, illustration, depth of field, great quality, amazing quality, very aesthetic, absurdres, cyberpunk, cyberpunk city,
Qwen Image Edit 2511
Edit an uploaded image with text instructions using Qwen Image Edit 2511. Upload a source image, replace the example instructions with the change you want, and run with the supplied defaults. Leave the negative prompt empty for a first run. The result is the decoded edited image. The uploaded image is automatically scaled and used both as a reference for prompt conditioning and as the initial image latent. There are no independent output-size controls. Denoise ranges from 0 to 1: lower values retain more initial structure, while 1 uses the full denoising schedule. Uploaded transparency is not preserved, and masked editing is not supported. Lightning mode switches together to the fixed four-step Lightning LoRA, Lightning steps and Lightning CFG. Standard mode defaults to 40 steps and CFG 4; Lightning mode defaults to 4 steps and CFG 1. Each mode's steps and CFG controls apply only while that mode is selected. Leave Lightning LoRA strength at 1 for the supplied configuration. Negative prompting has limited influence at CFG 1. Seed -1 selects new random starting noise on each run. Use a fixed seed to compare changes with the same starting noise. Sampler and scheduler apply in both modes; the supplied settings are Euler and simple. Advanced controls also expose guidance normalization strength, normalization placement and sampling shift, all shared by both modes. Normalizing before CFG can amplify noise; the default uses post-CFG attenuation only. Positive and negative reference-latent methods are separate controls for their respective conditioning paths. Keep their supplied index_timestep_zero values for a first run. These settings do not add more reference images or change output dimensions.
Two-pass image generator with tiled enlargement
Generate finished images from text with two Moody Z-Image diffusion passes and tiled enlargement, with optional face refinement and skin-contrast finishing. No uploaded image is required. The Pipe returns the finished image batch through Sweet Tea's image output; it does not promise permanent saving. For a first run, replace Positive prompt, review Negative prompt, and retain the 640 × 960 initial size, 1.7 Intermediate scale and 2× Enlargement factor. Image count determines the number of generated images. Final dimensions are approximately the initial dimensions multiplied by both scales, subject to latent-size rounding. Enlargement factor supports 0.05–4 in increments of 0.05; values below 1 reduce size. Larger dimensions and image batches increase processing and memory requirements. Positive prompt guides both generation passes and, by default, tiled enlargement. Negative prompt affects only the initial pass; subsequent stages use zeroed negative conditioning. Enable Use separate enlargement prompt to replace the main positive prompt during tile redraw only. Review the saved Enlargement prompt before using it because it contains portrait-specific styling. Enlargement denoise controls how much tile redraw can change the generated image. Keep the supplied sampling schedules and boundaries for a first run. Initially, a 17-step schedule runs from step 0 to step 12 and retains leftover noise. After latent enlargement, refinement uses an 11-step schedule starting at step 5 and ending beyond the schedule so it completes. Steps describe full schedule lengths, rather than simply executed iteration counts. Changing a schedule, its boundaries, scheduler or model timing changes the handoff or refinement behavior. Keep each starting boundary below its schedule length and stopping boundary. Each sampling stage has an independent seed; -1 requests fresh randomization. Fix all active stage seeds when comparing settings. Tile dimensions are calculated automatically from the decoded dimensions and Enlargement factor, using a fixed 96-pixel constant in the source formula. Changing Tile padding does not update that constant. Tile batch size controls how many tiles are processed together, rather than how many finished images are returned; it can change results and memory use. Seam fix mode defaults to None. Enabling a supported seam mode activates its separate redraw-strength, band-width, blur and padding controls. Sampler, scheduler and seam-mode choices without listed options depend on the installed node implementation. Automatic face refinement starts disabled. Enable its stage to detect and redraw faces, then replace the separate Face prompt for your subject; it does not inherit the main prompt, and its saved text is portrait-specific. Face denoise controls redraw strength. Detection threshold selects candidate faces, crop and mask controls determine the processed region, and guide and maximum sizes govern region-processing dimensions. Skin contrast also starts disabled. Enable its stage before adjusting Skin contrast strength. With normal blending, 0 keeps the incoming image and 1 uses the full model-processed result. When the stage is disabled, normal blending leaves the image unchanged; other blend modes can still alter tones by combining identical image branches. The diffusion checkpoints are fixed to the source's moody-wild-v4-undistilled-40steps-cfg4_00001__clean.safetensors and moody-real-v6.4_00001_.safetensors. Generic Z-Image dependency entries do not establish equivalent replacements. The retained native editor bridge and manual detailer remain hidden and bypassed; painted-mask refinement is unavailable in this interface.
Fashion editorial from three references
Create a fashion editorial image from separate character, background, and product references using Gemini. For a first run, upload all three required images and keep the example prompt. The references reach the model in this order: image 1 is the character and clothing, image 2 is the background, and image 3 is the product. The example requests a hard-flash, grainy editorial with the character showcasing the product. Edit the prompt's pose, expression, clothing, product placement, lighting, and film-style instructions to suit your references, then generate. Choose an explicit aspect ratio when the final layout matters; auto follows an input image's aspect ratio. Resolution defaults to 4K. The supported targets are 1K, 2K, and 4K, with Gemini's native upscaler used for 2K and 4K. Image model selects between the two source-supported Gemini models. Seed -1 randomizes each run; a fixed seed requests similar responses but does not guarantee repeatable results, especially when the model or settings change. Generation instructions set the provider's general behavior; keep their default initially and use Prompt for creative direction. Provider response can request IMAGE or IMAGE+TEXT. Only generated images are delivered by this Pipe; provider text is not connected to an output. References and prompts are sent to Gemini through a paid provider call. The result is a generative interpretation: exact identity, clothing, product details, and placement are not guaranteed. Uploaded alpha masks are not passed to generation, so uploaded transparency is not preserved by this workflow.
