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 41 758
Replace video background
Replace a video's background with a supplied image using Bria and receive one processed video. For a first run, upload the foreground video under Source video and a replacement image under Background image. Keep Background image enabled and leave File type and Codec on Auto, then run the Pipe. Background image is optional and bypassable. Bypassing it omits the replacement image; the supplied provider definition does not specify the resulting appearance. This Pipe has no replacement background video input. Uploaded image transparency is not carried through the connected RGB image output. Under Background replacement, Seed defaults to -1 to randomize on each run. Seed controls provider re-execution; results remain non-deterministic even with a fixed seed. Under Video output, File type offers Auto, MP4, MKV and WebM; Codec offers Auto, H.264 and AV1. Keep Auto for a first run, or choose a compatible container and codec together. The preserved hidden CRF control lacks documented semantics in the supplied evidence.
Subject and product composition
Generate a scene featuring a subject and product from two reference images using Gemini. Upload a subject image and a product image. In the scene prompt, image 1 means the subject and image 2 means the product. For a first run, keep the supplied prompt to request the subject holding the product with its logo visible, leave the other defaults in place, and run the Pipe. Edit the prompt to describe the interaction, setting, lighting, style and details to retain. Choose the output aspect ratio and resolution. Auto follows the reference aspect ratio; the product reference is resized and padded to match the subject image dimensions before both images are sent to Gemini. The 2K and 4K settings use Gemini's native upscaler. In Product reference preparation, padding selects white or black borders around the resized product, and interpolation selects its pixel resampling method. These controls affect the reference image rather than the generated scene background or final image size. Model selects either the supplied Gemini Pro image model or Nano Banana 2 (Gemini 3.1 Flash Image). Keep the supplied Generation instructions initially; edit them to change the provider's general behavior alongside the scene prompt. Provider response offers IMAGE or IMAGE+TEXT, but this Pipe returns only generated images through its image preview output; accompanying provider text is not returned. Use seed -1 for a new random seed each run. A fixed seed requests best-effort repeatability but does not guarantee identical results. Subject identity, product details and logo fidelity are requests rather than guarantees. Uploaded alpha masks are not connected, so uploaded transparency is not preserved through the reference path. Both reference images and the generation instructions are sent to Gemini for a provider-priced generation request.
Image grid
Combine two to four uploaded images into one collage. With all four inputs enabled and the default directions, the first pair forms the top row and the second pair forms the bottom row. For a first run, upload all four images and keep the defaults for a 2048 × 2048 output without gaps. Top-left and bottom-left images are required. Top-right and bottom-right are independently optional: disable their input stages when omitting them. With default directions, disabling both right inputs produces two images stacked vertically; disabling only one produces rows with different image counts. Open the advanced layout stages to change each pair's direction and the direction used to join the completed pairs. Each direction places the second image or pair relative to the first: right, down, left or up. Input position labels describe the default grid; changing directions changes their final positions. Size matching is enabled independently for each pair and for the final assembly. Gap widths and colors are also independent at those three joins. Gap widths are measured in pixels before the final resize, so resizing changes their final displayed size. A pair's direction, size matching and gap controls have no effect when its optional second image is disabled. Set the final width and height in pixels. Center cropping preserves proportions but can cut off collage content when its aspect ratio differs from the requested dimensions. Select disabled cropping to stretch instead, or set one dimension to zero to calculate it from the other while preserving aspect ratio. Keep at least one dimension greater than zero. Resampling controls pixel interpolation during this final resize. The result is one image returned through the Pipe output. Uploaded alpha masks are not connected, so uploaded transparency is not preserved.
Seedream 4 reference image generation
Create or edit images with Seedream 4 from one reference image and a text prompt, with an optional related-image mode. For a first run, upload your reference image and replace the example prompt with the result or edits you want. Describe which reference details should remain. Keep Output size at 2048x2048 (1:1) and Related images disabled to request one square image, then run the Pipe. This workflow requires one upload and does not provide additional reference inputs. Uploaded transparency is not preserved as an alpha mask. Choose an output size preset, or select Custom to use Width and Height in pixels. Custom width accepts even values from 1024 to 6240; custom height accepts even values from 1024 to 4992. These fields affect generation only when Custom is selected. For related scenes or variations, describe them in the prompt, set Related images to auto, and increase Maximum images. The model decides how many images to generate; Maximum images is a ceiling rather than an exact count. Input and generated images together cannot exceed 15, so with this one reference image keep Maximum images at 14 or below. Seed -1 randomizes each run; a fixed seed controls generation randomization. Enable Watermark to add an “AI generated” mark. Fail on partial results is enabled by default and aborts when requested images are missing or return an error; disable it to accept available partial results. Generation sends your reference image and prompt to ByteDance and uses a paid provider call.
ACE-Step song generator
Generate an MP3 song from music style tags and lyrics. Style tags describe the genre, instruments, vocals, tempo, and mood. Lyrics provide the sung text, with lyric strength available for tuning their influence. Choose the duration and number of variations; the default duration is 120 seconds. Advanced controls expose sampling, latent guidance, and MP3 encoding quality. Negative conditioning is automatically zeroed. For multilingual lyrics, the source notes require manual romanization and language markers such as [zh], [ja], or [ko].
Edit a photo and animate it
Edit a reference photo with FLUX Kontext, then animate the edited image with Wan 2.2. Each run returns an animated video and an edited starting-image preview. For a first run, upload a photo with a clearly visible subject. Replace or adapt Image prompt to describe the appearance, setting and visual style you want: the saved example requests a navy business suit in a modern office. Keep Motion prompt at its saved small-movement request, or describe the subject movement and camera behavior you want. Then run the Pipe and review both outputs. Image prompt controls the Kontext edit; Motion prompt controls the subsequent animation. In the advanced section, Video negative prompt discourages unwanted video content and movement. Its saved text combines Chinese visual-quality exclusions with English exclusions for fast movement, mouth movement, talking, visible teeth and camera zooms. Remove conflicting exclusions when requesting those actions. The negative prompt does not affect the image edit. Every run performs the image edit before animation. This interface has no separate animation-only or image-only mode. Video dimensions and seed come from the image-generation pipeline. Size, seed, duration, frame rate, sampling, Lightning acceleration, LoRA and enhancement settings remain at their saved workflow configurations and cannot be adjusted here. The workflow uses ComfyUI-UmeAiRT-Toolkit with the FLUX/Kontext and WAN_2.2/I2V model bundles. The broader source README describes additional configurations and toggles; those are not selectable modes in this Pipe.
Capybara: Image to Video
Create one silent video from an uploaded image using the fixed Capybara v0.1 model. The image supplies both the resized start frame and a separate visual reference. For a first run, upload an image and replace the example positive prompt with scene details, subject motion and camera movement appropriate to your image. Keep 640 × 640 pixels, 81 frames and 24 FPS for approximately 3.4 seconds of playback. Use positive dimensions divisible by 16 and frame counts of 4n + 1, such as 49 or 81. Larger dimensions and more frames increase generation work. The start image is resized to fit the output dimensions and padded when its aspect ratio differs. Padding color selects white or black borders; resize interpolation controls resampling. Reference crop selects center or none for the separate visual-reference encoding and does not change start-frame padding. Only the uploaded RGB image is connected; uploaded transparency is not preserved. Advanced controls include the negative prompt, seed, CFG, sampling settings and image preparation. Seed -1 randomizes each run; a fixed nonnegative seed reuses the initial noise. CFG controls the contrast between positive and negative conditioning. Begin with the existing sampling defaults: 20 steps, simple scheduler, Euler sampler, denoise 1 and model shift 7. Denoise changes the sampling schedule rather than directly controlling source-image preservation. Model shift changes the sampling model's noise-time behavior. FPS changes playback duration without changing the generated frame count. Choose 8-bit video for compatibility or 10-bit for smoother gradients where encoding and playback support it. File type and codec control the returned video's container and encoding; begin with auto for both, and use compatible combinations when selecting them manually. One video is returned through Sweet Tea's preview output, without a server-side save stage. No audio is connected. Batch size remains one in the normal interface because decoded frames are assembled into a single video rather than exported as separate variation clips.
Create a miniature architectural scene
Create an image of a miniature architectural model in a setting you describe. Upload a building photo or rendering; this Pipe generates a conceptual blueprint, then supplies both that blueprint and the original building image to the scene generator. It returns only the final scene image. Use the separate architectural blueprint tool for a standalone sheet. For a first run, upload your building image, edit the scene prompt to match its appearance, and keep the default 16:9 framing and 1K resolution. The supplied scene prompt requests a purple house on an architect’s desk with a blueprint displayed behind it. Describe the model, materials, surroundings, lighting, and camera view you want. Mention the blueprint in the scene prompt if you want it depicted. Scene aspect ratio and resolution control the final image. Auto framing asks the provider to match an input reference. Under Blueprint reference, edit the drawing views, annotations, and sheet layout for the intermediate image. Blueprint framing and resolution are independent of the final scene settings, but changes to that reference can influence the final composition. Both generation stages run on every request. Each stage has an independent model selector offering gemini-3-pro-image-preview and Nano Banana 2 (Gemini 3.1 Flash Image). Optional generation instructions guide each generator’s foundational behavior separately from its visual prompt; retain their defaults for a first run. Resolution choices are 1K, 2K, and 4K, with the provider’s native upscaler used for 2K and 4K. Each seed defaults to -1, which randomizes that stage on every run. Fixed seeds request best-effort consistency rather than identical results. Fixing only the scene seed does not stabilize an independently randomized blueprint reference. The blueprint and miniature scene are generated concept imagery. Geometry, dimensions, structural details, and annotations may be invented; use them for presentation rather than measured construction documentation. Higher resolution does not establish architectural accuracy.
Detail Tweaker LoRA (细节调整LoRA)
This is a LoRA for enhancing/diminishing detail while keeping the overall style/character; it works well with all kinds of base models (incl anime & realistic models)/style LoRA/character LoRA, etc.
Animate an image with Vidu Q1
Turn a starting image and motion prompt into a five-second, 1080p video with Vidu Q1. For a first run, upload your starting image, replace the example leopard prompt with action and camera instructions suited to that image, and leave Motion amount on auto. Generation uses the Vidu partner API and requires configured API access and sufficient credits. Motion amount requests the amplitude of object movement: auto lets the service choose, while small, medium, and large express an explicit preference. Describe camera movement and atmosphere in the prompt. Duration is fixed at five seconds and resolution at 1080p in this workflow; frame-rate wording in the prompt does not configure actual FPS. Seed -1 randomizes each run. A positive integer supplies a fixed generation seed without guaranteeing identical results; the provider treats 0 as random as well. The Video output advanced section offers automatic or explicit container and codec selection, plus CRF encoding quality. Leave File type and Codec on auto and Quality (CRF) on -1 for automatic settings. These controls affect output encoding rather than the generated action. The result is returned as a video through Sweet Tea.
Image to video with generated audio
Animate a starting image into a video with generated sound using prompted motion, a base pass, a fixed 2× latent upscale, and refinement. Upload a starting image and enter a prompt describing the scene, action, camera motion, and desired sounds or dialogue. The model, distilled LoRA, and upscaler are fixed. For a first run, retain the 720-pixel output long edge, 97 video frames, 97 audio frames, and 20 fps for conditioning, audio timing, and export. Keep MP4 file type and H.264 codec. The video plays for approximately 4.85 seconds. Output proportions follow the uploaded image; latent sizing can round dimensions. The long edge means the larger dimension, not a guaranteed height. Uploaded transparency is not retained through the RGB image path. When changing duration, set video and audio frame counts together. Video length uses an 8n+1 pattern, such as 97 or 121 frames. Keep conditioning, audio, and export frame rates equal, using a positive whole-number rate within all three fields' ranges. These independent controls do not update one another. Changing only export fps changes video playback duration without changing generated audio timing. Keep both batch sizes at 1 for the first run; if adjusting them, keep them equal. Advanced image guidance offers independent base and refinement strengths and switches to skip either image injection. A strength has no effect when that pass's injection is skipped. Skipping refinement image conditioning retains the upscaled base result and still runs refinement. With both injections skipped, generation is prompt-driven, but the upload remains required for aspect ratio and sizing. Reference resize and compression affect conditioning preparation rather than saved-video dimensions or encoding. Advanced sampling offers a shared seed and sampler, separate CFG values and noise schedules for each pass, and independent model and text-encoder strengths for the fixed LoRA. Retain source defaults initially. Seed -1 randomizes each run; a fixed nonnegative seed reuses noise initialization. The negative prompt feeds both passes, but standard CFG at the default value of 1 gives it no negative-guidance contribution. The base schedule contains eight intervals and the refinement schedule three; changing their comma-separated values changes sampling steps and noise levels. File type and codec are separate export controls. Select a compatible combination that supports audio to retain generated sound; GIF cannot carry sound. Export also exposes CRF, bitrate in Mbps, pixel format, bit depth, and color-space settings. Retain their defaults for the first run and keep codec, pixel format, and bit depth compatible when changing them. Selecting HDR does not establish HDR information in generated frames. Advanced decoding controls spatial tile size and overlap in pixels and temporal chunk size and overlap in frames; keep overlaps smaller than their corresponding sizes.
Grok image generation
Describe the image you want and generate one or more images with Grok. For a first run, use the supplied portrait prompt with the default model, square aspect ratio, 1K resolution, and one image. Change aspect ratio to set the composition and resolution to choose the output size tier; request 1–10 images for variations. The model selector offers the three Grok models present in the source. This workflow accepts text only and has no source-image editing input. Seed controls whether the provider node reruns, but the provider's results remain nondeterministic even with a fixed seed. The default seed randomizes on each run.
Hades style textual inversion
Textual Inversion trained on art from the Hades game, tested on Anything V3 model. Put this TI in the embeddings folder, type "Hades-8500" in prompt to activate. Also I recommend to use words "realistic", "cartoon",...
SD3.5 Large Canny-guided images
Generate new images from a prompt while following edges extracted from a reference image. Outputs include the generated image batch and one reference edge-map preview. For a first run, upload a reference image and replace the example prompt with your desired subject, colors, setting, and style. Keep output and reference dimensions at 1024 × 1024, Canny thresholds at 0.3 and 0.6, control strength at 0.66, and denoise at 1. The fixed SD3.5 Large checkpoint and matching Canny ControlNet are used together. Control strength sets how strongly reference edges influence generation. Set it to 0 for generation without edge influence; the reference upload is still required and its edge map is still returned. Guidance start and end are fractions of the denoising schedule from 0 to 1. Keep start at or below end. Reference resizing and generated output dimensions are independent. Match their aspect ratios for consistent framing. Center crop can remove content near the reference borders; disabled cropping resizes to the requested shape and can stretch the image. Reference interpolation changes resizing before edge detection, not generated output dimensions. Inspect the edge preview when changing crop, interpolation, or thresholds. Keep the low threshold below the high threshold; lower thresholds retain more faint detail, while higher thresholds select stronger contours. Variations sets the number of generated images using the same prompt and edge map. Sampling offers seed, steps, prompt guidance scale (CFG), sampler, scheduler, and denoise. Seed -1 selects a fresh random seed each run; use a fixed seed to compare settings with the same initial noise. Begin with the supplied sampling defaults. Denoise 1 uses the full generation schedule; lower values shorten denoising from the empty latent. The reference supplies edges rather than an initial image latent, so this tool does not preserve uploaded pixels. Uploaded transparency is not carried through the connected image path. Negative conditioning is zeroed internally; there is no editable negative prompt.
Anima image generation
Generate square images from text with Anima, using standard sampling or Turbo mode. No source image is required; the result is an image batch returned through Sweet Tea's preview transport, without permanent ComfyUI image saving. For a first run, replace the example positive prompt with your subject, composition, lighting and style. Review the negative prompt for attributes you want to exclude. Leave image count at 1 and Turbo mode disabled, then run with the supplied sampling defaults. Seed -1 produces fresh randomization on each run; use a fixed seed to compare changes with the same initial noise. Output size is fixed to the source's square, approximately 1-megapixel setting, about 1024 × 1024 pixels with dimensions rounded to multiples of 16. This interface has no resolution control. Image count determines the number of returned images; larger batches require more memory. Turbo mode selects the supplied Anima Turbo LoRA together with separate Turbo steps and guidance. Standard defaults are 30 steps and CFG 4; Turbo defaults are 8 steps and CFG 1. Only the active mode's steps and guidance affect sampling. Turbo LoRA strength affects the selected model only in Turbo mode. At CFG 1, negative conditioning does not influence ordinary classifier-free guidance. Advanced controls include both modes' steps and guidance, Turbo LoRA strength, seed, sampler, scheduler and denoise strength. Keep Euler, the simple scheduler and denoise strength 1 for the first run. Steps must be within 1–10000 and CFG within 0–100, even though their primitive input fields accept broader values. Lower denoise strength shortens the noise schedule; this text-to-image workflow has no source image to preserve. The runtime needs these source-specified model files in the corresponding ComfyUI model directories: [anima-base-v1.0.safetensors](https://huggingface.co/circlestone-labs/Anima/resolve/main/split_files/diffusion_models/anima-base-v1.0.safetensors) in models/diffusion_models; [anima-turbo-lora-v0.2.safetensors](https://huggingface.co/circlestone-labs/Anima-Official-LoRAs/resolve/main/anima-turbo-lora-v0.2.safetensors) in models/loras; [qwen_3_06b_base.safetensors](https://huggingface.co/circlestone-labs/Anima/resolve/main/split_files/text_encoders/qwen_3_06b_base.safetensors) in models/text_encoders; and [qwen_image_vae.safetensors](https://huggingface.co/circlestone-labs/Anima/resolve/main/split_files/vae/qwen_image_vae.safetensors) in models/vae. The selected output also requires [Sweet Tea nodes](https://github.com/tea-time-labs/sweet-tea-nodes). Standard and Turbo share the same image output.
Boogu Image Turbo
Generate an image or batch of images from a text prompt with the fixed Boogu Image Turbo model. No source image is required. For a first run, replace the example prompt with your subject, composition, lighting and style. Keep the square aspect ratio, 1-megapixel resolution and one image per run, then generate. The source sampling preset is 4 steps, CFG 1, lcm sampler, sgm_uniform scheduler and denoise 1. Aspect ratio and megapixels jointly determine the output dimensions. Megapixels measures total image resolution; the supplied size guidance describes a square 1 MP image as approximately 1024 × 1024 pixels. Images per run controls batch size. Larger resolutions and batches require more memory. Advanced Sampling exposes the seed, steps, prompt guidance, sampler, scheduler and denoise amount. Seed -1 selects a new random seed each run; use a fixed seed when comparing settings. Lower denoise values shorten the denoising schedule, but there is no source image whose structure can be preserved. Negative conditioning is zeroed internally, so there is no separate negative prompt. This tool does not provide image editing or model switching. Runtime requires boogu_image_turbo_fp8_scaled.safetensors in diffusion_models, qwen3vl_8b_fp8_scaled.safetensors in text_encoders, and ae.safetensors in vae, plus the Sweet Tea output node pack. The supplied dependency inventory marks these assets as missing; they must be available in the execution environment.
Chroma Radiance text to image
Create an image or batch of images from text using the fixed Chroma Radiance x0 model and T5 text encoder. No uploaded image is required; generated images are returned as the Pipe result. For a first run, replace the positive prompt with your subject, composition, lighting and desired style. Keep or edit the negative prompt to describe unwanted qualities, then run at the default 1024 × 1024 size, one image, 30 steps and CFG 3.5. The supplied miniature-bakers prompt is an example; its mention of 8K does not change output dimensions. Width and height set output dimensions in pixels and must be multiples of 16. Images per run sets the batch count. Larger dimensions and batches require more memory. Seed -1 randomizes each run; use a fixed nonnegative seed to reuse starting noise when comparing settings. Advanced controls include prompt guidance (CFG), steps, sampler, flow shift and two independent beta-schedule shape parameters. CFG controls the positive-versus-negative guidance strength. The scheduler is fixed to a beta schedule; Schedule alpha and Schedule beta jointly reshape its sampling-time distribution. The supplied sampling defaults are CFG 3.5, 30 steps, Euler, flow shift 1 and alpha and beta 0.4. Radiance rendering options control internal NeRF tiling and their active sigma range. NeRF tile size -1 uses the model default of 32; 0 disables tiling and may require substantially more memory. Radiance start sigma and Radiance end sigma specify the first and last sigma at which these options apply. Their defaults of 1 and 0 cover the full sampling range. Tile size does not determine output dimensions. The model and text encoder are fixed to chroma-radiance-x0.safetensors and t5xxl_fp8_e4m3fn_scaled.safetensors. This workflow generates images from scratch and provides no image editing, reference-image conditioning or upscale stage.
Photo to whole-body pose map
Upload a photo to produce a standalone pose map and an inspection preview blending that map with the resized photo and detected person boxes. Both outputs belong to the same pose extraction run; the preview is not a separate detection tool. For a first run, keep the 1024-pixel longer edge and the default drawing settings. The source default processes at most one person, choosing detections by descending confidence. Increase Maximum people for a group photo. Detection threshold controls which person boxes reach pose extraction; Keypoint threshold separately controls which estimated points are drawn. Lower thresholds can reveal uncertain detections or points, while higher thresholds can omit people or parts of a pose. The longer-edge setting drives both image resize branches and preserves aspect ratio. Choose which body regions to draw, then tune line width and face point size if needed. Preview blend and blend mode affect only the inspection preview, not the standalone pose map. This workflow produces rendered images, not a saved keypoint data file, and does not preserve uploaded transparency through its RGB image path. Required models are sdpose_wholebody_fp16.safetensors in checkpoints and rt_detr_v4-x-hgnet_fp16.safetensors in diffusion_models. Their supplied installation sources are https://huggingface.co/Comfy-Org/SDPose/resolve/main/checkpoints/sdpose_wholebody_fp16.safetensors and https://huggingface.co/Comfy-Org/SDPose/resolve/main/diffusion_models/rt_detr_v4-x-hgnet_fp16.safetensors.
Realisian Negative Embedding
This Negative Embedding is meant to be used with Realisian models but it can work on other realistic models (it has NOT been tested on anime models).
Corneo's Makima (Chainsaw Man) TI Embedding
This is a textual inversion embedding for Makima from Chainsaw Man. Chainsaw Man is an incredible series that you should try out if you haven't already.
