Ressourcenverzeichnis
Finden Sie, was Sie für Ihren Workflow benötigen.
Guides: techniques, setup, and troubleshootingDurchsuchen Sie Sweet Tea, Pipes, externe Workflows, Modelle, Knotenpakete und Anleitungen in einem faktenbasierten Katalog.
Beginnen Sie stattdessen mit einem Workflow ↗Suchergebnisse
Alle Ressourcen
20 angezeigt of 41.754
GPT Image edit and generation
Edit a reference image or generate images from text using GPT Image. For a first run, upload one 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: GPT Image 1.5, Auto output size, Auto background, High quality and one image. For text-only generation, disable the Source image stage and replace the editing instruction with a complete scene description. An upload is required while that stage is enabled. Model selects GPT Image 1, GPT Image 1.5 or GPT Image 2. Output size requests pixel dimensions or Auto. GPT Image 2 enables additional sizes and custom dimensions. To use custom dimensions, select GPT Image 2 and Custom output size, then enter width and height from 1024 to 3840 pixels in multiples of 16. These fields have no effect with other output-size selections. Background requests automatic, opaque or transparent background handling. Quality selects Low, Medium or High and affects generation time and cost. Image count requests 1 to 8 images using the same instruction and optional reference; the selected output returns those images. This tool connects one RGB reference image. Uploaded alpha transparency is not preserved, and the loader's separate alpha-mask output is unused. Although the generation node supports an optional inpainting mask, this tool does not connect or expose one, so it has no mask-based editing mode. The source seed setting is hidden because the backend does not implement it; it cannot make results reproducible.
fal.ai LTX Video
Describe a scene and its motion to generate a video through fal.ai's LTX Video 13B Distilled endpoint. This workflow accepts text only and returns the generated video. For a first run, enter a prompt describing the subject, setting, action, and camera movement. Keep the source defaults of 480p, 16:9, and 121 frames. The negative prompt provides a starting list of visual and motion defects to discourage. Resolution and aspect ratio set the video shape. Frame count sets its length in frames; this workflow does not expose FPS, so frame count is not a duration in seconds. Frame count must be entered as an integer, without units. Alternative shape values and frame counts must be supported by the provider. Seed defaults to randomization on each run. A fixed seed is available for exploring variations under otherwise similar settings, without a guarantee of identical provider results. Generation requires configured fal.ai access.
Qwen Image masked inpainting
Replace a masked region of an existing image with prompted content using Qwen Image and inpainting ControlNet. The result is composited over the working image; this tool edits within the existing canvas. For a first run, prepare an image whose longest edge is 1536 pixels and create its edit mask at that same size. Upload an alpha-bearing image with transparent pixels where replacement is wanted, or paint an edit mask if a loader mask editor is available. There is no separate mask-upload field. An opaque image without an edit mask provides no useful replacement region. Output transparency is not preserved. Replace the example replacement prompt with your desired content. Keep the longest edge at 1536, denoise at 1, steps at 20 and CFG at 2.5 for the initial run. Seed -1 randomizes each run; use a fixed seed to compare settings. A negative prompt can discourage unwanted content, but has little or no effect at CFG 1. Working image size resizes only the image, not its mask. Prepare the image at the chosen size before masking, or disable that stage to retain the uploaded dimensions. Changing the resize target after masking can misalign the edit region. Image resampling affects only this resize operation and is inactive when the stage is disabled. Mask expansion grows or contracts the editable region in uploaded-mask pixels; tapered corners affect nonzero expansion. Blur radius sets the Gaussian filter's support, and sigma independently controls its spread. The processed mask controls latent noise, inpainting guidance and the final blend. Pixels outside that mask come from the working image, which may be resized; softened edges blend original and generated pixels. Lower denoise retains more existing content within the edit region. Inpainting guidance strength controls the influence of the image and mask, while Guidance start and Guidance end set its active sampling interval as fractions from 0 to 1. Keep start below end. Sampler and scheduler select the denoising algorithm and noise-level distribution; retain Euler and Simple for the first run. Sampling shift is a separate model-specific schedule adjustment with source default 3.1. The optional Lightning LoRA starts disabled. When enabling it at strength 1, try 4 steps and CFG 1 together, as described in the source guidance. Its toggle does not change those independent controls. When disabling Lightning, manually restore 20 steps and CFG 2.5 to return to the baseline configuration. For canvas expansion with directional padding and an automatically created padding mask, use the separate outpainting tool.
Seedance text to video
Generate one short video from a scene description with Seedance, with optional generated audio on Seedance 1.5 Pro. No uploaded media is required. For a first run, keep Seedance 1.5 Pro selected and use the example prompt or describe your own subject, setting, action, and camera movement. The defaults request a five-second, 720p, widescreen video with generated audio. Include dialogue or sound intent in the prompt when audio is enabled, and keep the action and dialogue appropriate for the chosen duration. Choose a duration from 3–12 seconds, a resolution of 480p, 720p, or 1080p, and the aspect ratio for your intended frame shape. Model lets you select the listed Seedance 1.0 alternatives; those models ignore Generate audio. With Seedance 1.5 Pro, turn Generate audio off for a silent result. Fixed camera adds a stationary-camera instruction to the prompt but does not guarantee camera lock. Remove conflicting camera movement instructions when enabling it. AI watermark requests an “AI generated” watermark. Seed -1 randomizes each run; a nonnegative seed controls the generation's random input. The generated video is returned through Sweet Tea's video output transport with automatic format and codec selection by default. Generation sends your prompt to the ByteDance provider and uses provider-priced billing.
Amateur Snapshot Photo - [STYLE] [LORA] [FLUX]
Latest page update: Updated recommended workflow again to a version 2, which is slightly even better.
Grok reference-to-video
Generate a short Grok video guided by two reference images and a description of the scene and motion. Upload both reference images, then describe the action, visual style, camera movement, and intended role of each image. For a first run with the supplied prompt, use a smartphone image as reference 1 and a character image as reference 2. For other subjects, replace the example prompt to match your uploads. Both images guide content or style; they are not designated start and end frames. The default requests a six-second, 720p landscape video at 16:9. Choose a duration from 2 to 10 seconds in whole-second increments, one of the available landscape, square, or portrait aspect ratios, and 480p or 720p resolution. Describe how the two references should relate within the chosen duration. Leave Seed at -1 for a fresh value each run. A fixed seed controls provider re-running and caching, but generated results remain nondeterministic. Under Video encoding, leave file type and codec on auto for automatic selection, or choose the available container and codec settings. Generation uses the hosted Grok service through ComfyUI's partner API and requires configured access. Reference images guide the result rather than guarantee exact reproduction.
bad X (bad sdxl negative embedding)
This is a textual inversion based on SDXL 1.0 images with the refiner, tested and trained to enhance the effectiveness of the output of human figures. It is meant to correct mutation like symptoms of the face and...
Wan 2.1 VACE text to video
Generate a silent video from text using Wan 2.1 VACE 14B. For a first run, replace the supplied car-scene prompt with your subject, setting, visual style and intended motion. Keep the remaining defaults: 1280 × 720 pixels, 81 frames, 16 FPS, batch size 1, CausVid enabled, 4 sampling steps and CFG 1. The output lasts about 5.1 seconds. No media upload is required. Width and height use multiples of 16. Frame count advances by four from 1. Frame count determines how many frames are generated; FPS changes playback speed and duration without adding frames. Keep batch size 1 for a single sequence: this Pipe assembles one video output and does not offer separate files per batch item. The advanced negative prompt discourages unwanted traits; its supplied Chinese text includes static frames, blur, overexposure, text overlays and visual artifacts. CausVid uses a fixed LoRA with separate model and text-encoder strengths. The supplied guidance recommends testing model strength between 0.3 and 0.7 if motion is shaky or blurry. To use the base-model mode, bypass CausVid and manually set Sampling steps to 20 and CFG to 6. Bypassing does not update these settings automatically. Use seed -1 for randomization or a fixed nonnegative seed to reuse starting noise. Sampler, scheduler, denoising amount and schedule shift remain available for advanced adjustment; keep their supplied defaults for a first run. Video playback includes an 8- or 10-bit choice; 10-bit may reduce gradient banding but is not supported by every player or downstream node. Video export provides container, codec and CRF controls. Keep automatic container and codec selection and CRF -1 for the first run; these settings affect encoding rather than generated content. This is a text-only configuration: no reference image, control video, control mask or audio is connected. The VACE strength field does not add those inputs. The model and CausVid LoRA are fixed; smaller-model alternatives and animated WebP export are not selectable modes in this Pipe.
Qwen-Image edge-guided generation
Upload a source image and describe the image you want to create. The Pipe resizes the upload to approximately the chosen megapixel count while retaining its aspect ratio, extracts Canny edges, and uses the encoded edge map as both reference conditioning and the sampler's starting latent. It returns the generated image and the edge-map preview. For a first run, keep the 1 MP size, default edge thresholds, full denoise, and standard mode. Replace the example positive prompt with your desired subject, appearance, and style. Inspect the edge preview: lower thresholds admit more edges, while higher thresholds suppress weaker details. Keep the low threshold below the high threshold. Enable Lightning mode to select the supplied four-step Lightning LoRA together with its separate steps and CFG settings. Standard mode defaults to 20 steps and CFG 2.5; Lightning defaults to 4 steps and CFG 1. Each mode retains its own tuning fields, and only the selected mode's fields affect a run. Denoise operates on the edge-map latent, not the original photograph. Lower values can leave more of that edge-map appearance in the result; this is not a control for preserving the source image's colors. Union LoRA strength adjusts the model adaptation used for reference guidance rather than guaranteeing an exact edge match. Keep Lightning strength at 1 initially. The resize stage can be disabled to work at the upload's native dimensions; megapixel and resampling settings then have no effect, and large uploads can substantially increase resource use. This graph implements Canny guidance only, despite the Union model supporting other control types. It does not use the upload's alpha mask.
Flux.1 Inpaint
Fill missing parts of images using Flux.1 Fill Inpainting.
Magnific reference style transfer
Create a restyled image from a source image, a visual style reference and a descriptive prompt. For a first run, upload both images and replace the fantasy example prompt with a description of your subject and intended appearance. Keep style strength 70, structure strength 41, faithful flavor, balanced engine and fixed generation enabled, then run the Pipe to view the resulting image. Style strength controls the reference style's influence from 0 to 100 percent. Structure strength controls retention of the original image's structure on a separate 0–100 scale. Adjust these independently to balance restyling with composition retention. Flavor selects the service's style-transfer treatment, while engine selects its processing engine. Both offer alternative treatments of the same inputs. Disable fixed generation to introduce more randomness across runs; enabling it does not guarantee identical results. Processing sends the images and prompt to the Magnific partner service through its API node, requires service authentication and uses provider-priced generation. Portrait enhancements are unavailable in this Pipe and remain disabled. Uploaded alpha masks are not connected, so source transparency is not preserved.
Gemini image refinement
Refine and upscale a reference image with Gemini, or generate a new image from a prompt. For a first run, upload a Source image, leave that stage enabled, and use the default prompt to request refined details, textures and materials while retaining composition and text. The default resolution is 4K. This is generative editing: exact text, details and composition may change. Uploaded transparency is not preserved because only the loader's IMAGE output is connected. Change Prompt to describe other edits and what should remain. Aspect ratio set to auto follows the reference image's proportions; an explicit ratio requests a different output shape. Resolution selects a target size tier of 1K, 2K or 4K rather than exact pixel dimensions. The 2K and 4K settings use Gemini's native upscaler, and dimensions also depend on aspect ratio. For text-to-image generation, bypass Source image and replace the refinement prompt with a complete description of the image you want. Choose an explicit aspect ratio to specify the output proportions. This Pipe accepts one reference image. Model selects the Gemini image model. Seed -1 requests a new random seed each run; a fixed seed requests best-effort repeatability, but identical results are not guaranteed. System instructions set foundational provider behavior alongside Prompt; the default directs Gemini to produce an image. Provider response format requests either IMAGE or IMAGE+TEXT, but only the image is connected to this Pipe's result. Generation uses the Gemini provider and sends the prompt and any reference image to that service through a paid provider call. The generated image is returned through Sweet Tea's preview transport; this workflow does not permanently save an image file.
MoGe valid geometry mask
Estimate a geometry-validity mask from one perspective image. For a first run, upload an image, leave Automatic resizing enabled and keep Detail level at 2. Each run performs MoGe geometry inference, renders predicted validity and converts the rendered red channel into the output mask. The mask identifies regions with valid predicted geometry. It is not foreground segmentation or an object cutout; sky and other regions with undefined geometry may be excluded. Prediction uses the uploaded RGB image, independently of uploaded transparency. This Pipe exports a mask, not depth, normal maps or a mesh. Automatic resizing checks image width. Images wider than 2048 pixels are resized to a fixed 2048-pixel longer edge before inference. Images at most 2048 pixels wide remain unchanged, even when their height exceeds 2048. Disable Automatic resizing to use the original image. The mask follows the dimensions used for inference, so resizing can change its dimensions relative to the upload. Open Geometry estimation for inference controls. Detail level ranges from 0 to 9; the source identifies 0 as fastest and 9 as most detailed. Horizontal field of view is measured in degrees: leave 0 for automatic recovery from predicted points, or enter a known camera value for geometry unprojection. Exclude invalid geometry sets invalid points and depth to infinity when enabled; disabling it retains raw predicted geometry. The validity mask is returned separately in either mode. Inference batch size controls images per inference call and does not create additional variants from the single upload. Force projection starts enabled; its precise effect is not documented in the supplied source, so retain the default for a first run. These settings do not select other output types. Mask rendering and red-channel conversion are fixed to preserve the validity-mask result. The workflow uses the fixed model moge_2_vitl_normal_fp16.safetensors. For a manually managed ComfyUI installation, the source specifies models/geometry_estimation/ as its location and https://huggingface.co/Comfy-Org/MoGe/resolve/main/geometry_estimation/moge_2_vitl_normal_fp16.safetensors as its download source.
Princess Style
This is a Style embedding which turns your image subjects into bonafide Princesses! Think pearls, tiaras, and ballgowns! Simply throw Style-Princess on to the end of your prompt (or the start, if the effect doesn't...
Z Image Turbo pixel-art image generator
Generate one image or a batch from text using the fixed Z Image Turbo model, with optional pixel-art styling. No uploaded image is required. For a first run, leave pixel-art style enabled, edit the pixel-art prompt, and keep the 1024 × 1024 size, batch size 1, LoRA strength 1 and four sampling steps. The source prompt begins with “Pixel art style, retro”; use that as a starting point when describing your own subject. Run the Pipe to receive the generated images. The style toggle switches both the model branch and the prompt. When disabled, the base model uses the separate unstyled prompt. Edit both prompts when comparing modes; changes to one do not update the other. LoRA strength affects only the enabled pixel-art branch. Width and height are output pixels and must be multiples of 16. Batch size controls how many images are returned. The output stage does not permanently save images to the ComfyUI output folder. Advanced sampling controls include seed, steps, guidance scale, sampler, scheduler and denoise fraction. Seed -1 randomizes each run; use a fixed seed to compare changes with otherwise unchanged settings. The source defaults are four steps, guidance 1, res_multistep sampling, the simple scheduler and denoise fraction 1. Negative conditioning is zeroed automatically, so there is no editable negative prompt. The optional sampling-shift stage starts bypassed, matching the source workflow. Enable it before adjusting its shift value; the source value is 3. While bypassed, shift has no effect. Generation starts from an empty latent and does not support image-to-image input. Reducing denoise fraction changes sampling but does not preserve an uploaded image. The model, text encoder, VAE and pixel-art LoRA are fixed by this Pipe.
Video to pose animation
Turn a source video into a rendered whole-body pose animation using SDPose. The output is one pose video, rather than an overlay on the original footage or a downloadable keypoint dataset. It uses the source frame rate and includes source audio when available. For a first run, upload a clip with one clearly visible person, keep the default 1024-pixel longer edge and drawing settings, and run the Pipe. The required checkpoint is sdpose_wholebody_fp16.safetensors. Frames are resized while preserving aspect ratio before pose extraction; Longer edge sets both detection resolution and rendered frame dimensions. Use Draw body, Draw hands, Draw face, Draw feet and Draw head to choose the rendered features. These switches change visualization, not what the model extracts. Confidence threshold controls which keypoints are drawn: higher values suppress uncertain points, while lower values include more points. Line width and Face point size are measured in output pixels; Face point size applies when Draw face is enabled. Resampling method controls frame resizing before detection. Area is suited to downscaling and Lanczos to upscaling. Processing batch size controls how many frames the extractor processes together; lower it if memory is limited. It does not change video duration. Frame rate is inherited from the source and cannot be independently adjusted here. This workflow supplies no bounding boxes, which the extractor requires for multi-person detection. Use single-person footage rather than expecting reliable multi-person capture. Advanced output controls select the container, codec, encoding quality and video bit depth. Keep File type and Codec on automatic, Quality (CRF) at -1 and Bit depth at 8 for a first run. Explicit container and codec choices must be compatible. The 10-bit option requires compatible encoding and playback and does not improve pose detection.
OmniGen2: Text to Image
Generate a single image or a batch from text with OmniGen2. No uploaded image is required; the diffusion model, Qwen text encoder, and VAE are fixed by this workflow. For a first run, enter the subject, composition, lighting, and style in Positive prompt. Negative prompt describes unwanted features. Leave the sampling defaults in place and generate one 1024 × 1024 image. Width and height are measured in pixels and must be multiples of 16; together they determine the aspect ratio. Images per run sets the batch size. Larger dimensions and batches increase memory use. Advanced settings expose guidance, seed, and sampling controls. Guidance defaults to 5. Guidance mode selects the source dual guider's regular or nested combination. Negative prompt feeds both its secondary and negative branches; there is no separate secondary prompt. Secondary guidance defaults to 2 and is not an independent negative-prompt strength in regular mode. Seed -1 chooses new random noise for each run. Use a fixed nonnegative seed when comparing settings. Sampling defaults are 20 steps, Euler sampler, and the simple scheduler. Steps sets the sampling step count, Sampler chooses the sampling algorithm, and Scheduler sets the distribution of noise levels. Denoise selects the fraction of the schedule used and defaults to 1 for the full schedule. Sampling begins from an empty latent. Lowering Denoise does not enable image editing, and this tool has no reference-image input. All decoded images in the batch are returned through Sweet Tea's preview output. The workflow requires omnigen2_fp16.safetensors, qwen_2.5_vl_fp16.safetensors, and ae.safetensors; model selection is not exposed.
LTX-2.5 text to video with audio
Create one video with generated audio from a text prompt using the fixed LTX-2.5 distilled model. Describe subjects, action, camera movement, lighting and desired sounds in Positive prompt. No image or audio upload is required. For a first run, replace the example prompt or try it as supplied, then keep 5 seconds, 24 FPS, widescreen aspect ratio and the 0.9-megapixel resolution target. Leave output file type and codec on auto. Prompt enhancement is enabled: it expands your text before video conditioning. Disable Enhance prompt to encode your original wording directly. Expansion tuning affects the output only when enhancement is enabled; sampling seed, temperature and token filters apply when Prompt sampling is on. The pipeline jointly generates audio and video at half the selected spatial dimensions, upscales the video latent with the fixed 2× spatial upscaler, then refines audio and video together. Both streams are decoded and assembled into the returned video. Upscaling and refinement are integrated stages, with no bypass controls. Aspect ratio and megapixels jointly determine approximate final dimensions through the resolution selector. They are shape and pixel-budget choices, not separate width and height overrides. Duration is in positive whole seconds. Use a whole-number frame rate from 1 to 120 FPS. The graph requests duration × FPS + 1 frames: keep duration × FPS divisible by eight and no greater than 999, respecting the audio latent stage's 1,000-frame maximum. At 24 FPS, whole-second durations from 1 to 41 meet these frame-count requirements. Playback includes the extra frame. Larger dimensions and more frames increase processing and memory requirements. Advanced controls separate initial generation from refinement. Each stage has its own seed, video guidance, audio guidance, sampler and comma-separated noise schedule. Keep the supplied guidance value of 1, samplers and distilled-model schedules for the first run. Negative prompt supplies unwanted-content conditioning to both stages; its influence depends on their guidance settings. A seed of -1 randomizes each run. To hold the exposed random inputs constant, fix both stage seeds and, when sampled prompt enhancement is enabled, its separate seed. Fixing only Generation seed leaves other random inputs variable. Video decode tiling controls spatial tile sizes and overlap in pixels, plus temporal chunk sizes and overlap in frames. It does not set output resolution or duration. Video output encoding offers automatic settings or explicit container, codec and CRF choices; use compatible container and codec selections and keep CRF at -1 for the transport default. The runtime needs the fixed transformer, generation text encoder, enhancement text model, audio and video VAEs, and latent upscaler named in the workflow. This tool returns video with generated audio and does not offer uploaded audio, a silent-output switch, independent audio export or automatic duration prediction. Audio and video batch fields remain at one for this single-video interface.
SDXL image remix
Remix an uploaded image using a text prompt and the fixed waiNSFWIllustrious_v130.safetensors checkpoint, with optional LoRAs, tiled enlargement and subject refinement. The saved result includes every enabled finishing stage. For a first run, upload a source image, replace the example positive prompt with your desired subject, appearance and style, and review the negative prompt for unwanted content. Keep automatic resizing enabled and the default 1024 × 1024 generation dimensions. Leave LoRAs, upscaling and subject refinement off initially. Denoise starts at 0.75: lower values retain more of the source, while higher values allow stronger reinterpretation. Width and height apply to base generation when automatic resizing is enabled; source-size handling with resizing disabled is determined by the toolkit. Seed -1 requests a fresh random seed each run. Use a fixed seed when comparing prompts or settings. Steps, CFG, sampler and scheduler control base sampling. The checkpoint and baked VAE are fixed; CLIP skip is available under advanced model settings. Each of the three LoRA slots has an independent enable switch, installed LoRA name and strength. Select a LoRA compatible with the fixed Illustrious checkpoint and enable its slot. Multiple enabled slots combine in the same generation. Optional upscaling runs after base generation and before subject refinement. It uses the fixed 4x_NMKD-Siax_200k.pth model; the upscale factor multiplies the base image dimensions. Upscale denoise controls changes during enlargement. Tile width, height and padding are in pixels. Seam correction starts at None; its related settings apply when a supported seam mode is selected. Prompt cleanup, tile-processing mode, uniform tiles and tiled decoding are also available. Upscale sampler and scheduler start at Pipeline. Upscale steps retains the source value 0, whose exact fallback behavior is not established here; retain it initially and consult the installed toolkit before changing it. Optional subject refinement runs last, targets face by default and has its own denoise control. Enter only a subject category supported by the installed toolkit. Its advanced detection, crop, feather and mask controls preserve the source defaults. Several have undocumented types or units, including boolean guide size, integer noise mask 5 and detection threshold 1024. Consult the matching UmeAiRT Toolkit documentation before changing these values; no normalized confidence range or pixel interpretation is established for them. Requires ComfyUI-UmeAiRT-Toolkit (https://gitlab.com/UmeAiRT-Studio/ComfyUI-UmeAiRT-Toolkit.git) and waiNSFWIllustrious_v130.safetensors in ComfyUI/models/checkpoints/. Enabling upscaling additionally requires 4x_NMKD-Siax_200k.pth in ComfyUI/models/upscale_models/. Optional LoRAs belong in ComfyUI/models/loras/. The source describes subject-refinement detector and segmentation assets under models/ultralytics/bbox/, models/sams/ and models/ultralytics/segm/, with automatic downloads handled by the toolkit; exact asset names are not supplied. This Pipe requires an uploaded image and exposes one final remixed-image output. It does not expose checkpoint replacement, a separate text-to-image mode or standalone finishing of an upload without upstream generation.
Seedance image to video
Animate a starting image into one video using a motion prompt and ByteDance Seedance. Upload the image to use as the first frame, then describe the subject's action and desired camera movement. Generation sends the image and prompt to the ByteDance API and requires configured API access and sufficient credits. The result is returned through Sweet Tea's video output. For a first run, replace the example prompt to match your image and keep Seedance 1.0 Pro, 480p, adaptive aspect ratio and a duration of 5 seconds. Duration accepts whole seconds from 3 to 12. Resolution and aspect ratio set the requested output shape. Choose among the available Seedance models. Generate audio applies only to seedance-1-5-pro-251215; other models ignore it. Fixed camera adds a stationary-camera instruction to the prompt without guaranteeing the effect, so keep your motion prompt consistent with that request. Seed -1 randomizes each run; a specified seed controls generation randomness. AI watermark requests an “AI generated” mark on the video. Advanced Video output settings select the returned file container and codec. Leave both on auto and Quality (CRF) on -1 for default output encoding, or choose from the exposed encoding controls.
