Resource directory
Find what you need for your workflow.
Guides: techniques, setup, and troubleshootingSearch Sweet Tea Pipes, external workflows, models, node packs, and guides in one evidence-linked catalog.
Start with a workflow instead ↗Search results
All resources
19 shown of 41,749
Kling motion transfer
Animate a reference character image using movement and expression from a motion reference video, producing one video through Kling motion control. For a first run, upload a character image and a 3–30-second motion clip, preferably with similar aspect ratios. Leave Character orientation set to video, Keep original sound enabled, generation mode set to pro and model set to kling-v2-6. Replace the example studio-setting prompt with your desired setting, lighting and visual details, then run the Pipe. Character orientation determines which reference supplies the character's facing direction. In video mode, movement, expression, camera motion and orientation follow the reference video; the clip must be 3–30 seconds. In image mode, movement and expression still follow the video, while orientation follows the image and camera details come from the prompt; the clip must be 3–10 seconds. Prepare a clip within the selected mode's duration limit before uploading. Disable Keep original sound if you do not want reference-video audio retained. Generation mode offers pro and std, and Kling model offers kling-v2-6 and kling-v3. The supplied source does not explain their comparative behavior. This stage sends the references and prompt to the Kling provider and uses a provider-priced generation call. The advanced Video output section provides file type, codec and CRF controls. Keep auto for file type and codec and -1 for CRF on a first run. The supplied evidence does not specify container/codec compatibility or codec-specific CRF behavior. The selected output returns one video through Sweet Tea rather than using the original workflow's SaveVideo stage.
Topaz Bloom 2 image enhancement
Enhance an uploaded image with the workflow's fixed Topaz Bloom 2 preset and preview the enhanced result. For a first run, upload your source image and run the Pipe. The image upload is required and is the only editable control. The source configuration uses an empty prompt; its accompanying guidance says this lets the model create a prompt automatically. Prompt, seed, model selection and enhancement settings are fixed in this Pipe. No alternate modes or optional stages are exposed. Uploaded transparency is not preserved. Output scale and repeatable results are not established by the supplied source evidence. The selected output is an image preview; the original save node's PNG encoding settings do not apply to this output.
Create images from two references
Generate or edit an image with Gemini using two reference images and a written instruction. Upload both references, then explain which subjects, details or style to take from each. Both uploads are required and are sent together as context for one image request, rather than processed into separate outputs. For a first run, upload character and clothing references, describe how to combine them in a casual fashion pose, and keep the supplied 1:1 aspect ratio and 1K resolution. The references have no fixed character or clothing roles; you can request other compositions or edits. Choose an explicit aspect ratio for the output shape. Auto asks Gemini to match the input image proportions; an explicit ratio avoids ambiguity when the references differ. Resolution selects 1K, 2K or 4K; 2K and 4K use Gemini's native upscaler. The model selector offers Gemini 3 Pro Image Preview and Nano Banana 2 (Gemini 3.1 Flash Image). System instructions set foundational behavior; the supplied instructions prioritize interpreting the prompt visually and producing an image. Keep IMAGE for image-only responses, or select IMAGE+TEXT to request accompanying text from Gemini. This Pipe returns only generated images, so accompanying text is not delivered. Use seed -1 for a new random seed each run. Fixed seeds request best-effort consistency, not identical results. Reference details are interpreted by the model, so exact identity or garment reproduction is not guaranteed. Uploaded alpha masks are not connected, so reference transparency is not preserved. This workflow sends the prompt and reference images to Gemini through a paid provider request. Generated images are returned to Sweet Tea through an output transport that does not permanently save them through the workflow.
Gemini image editing with document context
Create an image from a required reference image and written instructions using Gemini, with optional text or PDF context. The generated image is delivered as the Pipe result. For a first run, upload a reference image, replace the example prompt with the changes you want, and describe which details to retain. Leave Aspect ratio on auto to follow the reference image's proportions, leave both document stages disabled, and run the Pipe. Choose an explicit aspect ratio to request different output proportions. To add document context, place a .txt or .pdf file in ComfyUI/input, enter its filename in the relevant document field, and enable that stage. These fields select existing files rather than uploading documents, and their empty defaults must be filled before enabling them. Either stage can be used independently; enable both with distinct files to supply two documents. Explain in your prompt how Gemini should use their contents. Seed -1 randomizes each run. A fixed seed requests best-effort consistency; identical results are not guaranteed. Generation instructions govern Gemini's interpretation of the prompt. Retain the directive to produce an image when editing them. Response format supports IMAGE and IMAGE+TEXT, but accompanying text is not delivered by this Pipe. This workflow requires a reference image even though Gemini supports other generation modes. Uploaded transparency is not preserved through the connected RGB image path. Image creation uses the Gemini provider and sends the reference image, instructions and enabled document context to that provider; provider billing applies.
Wan 2.1 video outpainting
Expand an uploaded video's canvas with generated surroundings using Wan 2.1 VACE, and export an H.264 MP4 with source audio when available. For a first run, upload a continuous scene and enter an outpainting prompt describing both the existing scene and the surroundings to add. Keep both aspect-ratio controls on Horizontal and retain the other defaults. If you change the canvas shape, set Aspect ratio and Continuation aspect ratio to the same choice: their generated batches must have matching dimensions. Both stages use the fixed I2V720p/HQ resolution preset. The input is resampled to 24 frames per second. Initial source size and Continuation source size set the original content's size in megapixels before padding to the target canvas. Smaller source frames can leave more space for generated surroundings; use matching sizes across stages for consistent framing. Each stage has its own mask-feathering and VACE control-strength settings. Control strength adjusts guidance from the padded source frames and masks. The workflow generates an initial 81-frame batch followed by three fixed continuation iterations. Each continuation requests 81 generated frames, carries the previous generated batch's last frame as a reference, and advances the source slice by 81 frames. The continuation source slices request up to frame index 405, approximately 16.9 seconds at 24 FPS. Supply enough continuous footage to cover those requests. This tool does not adapt its iteration count to the uploaded clip's length. The source's value labeled Length in Seconds controls loop arithmetic rather than exact playback duration and is not an editable duration control. Continuation slices request more conditioning frames than each generated batch; the supplied evidence does not establish precisely how excess conditioning frames are handled. Both stages share the prompt, negative prompt, seed, and CausVid LoRA. Seed -1 randomizes each run; a fixed seed reuses the shared noise seed. Initial and continuation sampling controls are independent. The supplied recipe uses CausVid strength 1, three Euler steps, normal scheduling, CFG 1, and denoise 1 in both stages. Keep these defaults for the first run; changes to LoRA strength affect both stages. Compression controls H.264 CRF: lower values retain more detail and generally create larger files. Trim to audio is enabled by default and can shorten the encoded video to the source audio's length. Disable it to retain generated frames beyond shorter audio. Playback remains fixed at 24 FPS. Outpainting synthesizes content and can alter the original region or show transitions between batches. Scene cuts can conflict with the previous-frame continuation reference. That reference remains part of the workflow and has no supported disable switch.
Boogu 0.1 image edit · Int8
Create edited images from one uploaded reference and natural-language instructions using the fixed Boogu Image Edit convrot Int8 model. Describe both the changes you want and anything that should stay unchanged, such as the subject, clothing or layout. The results appear as an edited-image batch in Sweet Tea. For a first run, upload your source image, replace the example editing instruction, leave Negative prompt blank, and keep Output megapixels at 1, Variants at 1 and the default sampling settings. Negative prompt describes features to discourage. Output size measures a resized copy of the source to derive output dimensions at the requested megapixel count while retaining its aspect ratio. The original uploaded image remains the editing reference. Bypass Output size to derive dimensions directly from the uploaded image instead; Output megapixels then has no effect. Final dimensions may be rounded by latent sizing. Variants controls how many edited images are generated from the same reference and instruction. Larger resolutions and batches require more memory. Seed -1 randomizes each run; use a fixed seed to reuse the initial noise with otherwise unchanged settings. Advanced controls include CFG, steps, sampler and scheduler. Sampling schedule fraction uses only part of the noise schedule when below 1; it is not an image-preservation strength. Keep Add noise enabled and the schedule fraction at 1 for the intended operation. Schedule shift affects the scheduler's noise-level calculation, while the sampler receives the original loaded model. Keep the source defaults initially. This tool generates from an empty latent conditioned on the reference, rather than directly denoising the uploaded pixels. Exact identity and layout preservation are not guaranteed. Only the loader's image output is connected; the uploaded alpha mask is unused, so uploaded transparency is not preserved. The selected output uses Sweet Tea preview transport rather than the original file-saving stage; the original PNG export settings do not apply.
Qwen Image Layer Decomposition
Turn a source image into a batch of generated layer images, guided by an overall image description. Upload your image and replace the example description “guitar” with a description of its objects, appearance and arrangement, including relevant partly occluded elements. The description guides the whole decomposition; there are no individual layer prompts or controls assigning objects to particular layers. For a first run, keep the maximum input dimension at 640 pixels, request two layers and use one decomposition per run. Leave the negative prompt empty unless you want to discourage particular attributes. Keep the sampling defaults: 20 steps, CFG 2.5, Euler sampler, simple scheduler, denoise 1 and sampling shift 1. The maximum input dimension sets the reference image’s longest dimension in pixels; the other dimension follows its aspect ratio. Output dimensions derive from this resized reference rather than independent width and height controls. Resize method changes the interpolation applied before reference encoding. More layers, larger images or more decompositions per run increase processing demands. Multiple decompositions use the same source and description, and their generated layers are returned together as an image batch. Use seed -1 for randomization on each run, or a fixed seed to compare settings with the same sampling noise. CFG controls text-conditioning influence. Sampler and scheduler select the denoising algorithm and noise schedule; sampling shift separately adjusts the model sampling schedule. Denoise controls sampling from an empty layered latent, so lowering it is not a source-image preservation strength. Results are generated image layers rather than an editable layered document. Sweet Tea returns the output batch through its image preview transport; this workflow does not permanently save images through ComfyUI SaveImage. The source loader’s RGB image output supplies the reference; uploaded alpha masks are not connected or preserved as input masks. The model, text encoder and layered VAE remain fixed to the source configuration.
ACE-Step music remix
Remix an uploaded music track with ACE-Step using style tags and lyrics, then export the result as MP3. For a first run, upload the audio excerpt you want to remix, replace the example style tags and lyrics with your intended direction, and keep the sampling defaults, including remix strength 0.3. Describe genre, instruments, mood, tempo and vocal character with comma-separated tags. Enter intended sung words with section labels and line breaks. Lyrics and tags guide generation; they do not guarantee exact words or performance. Remix strength ranges from 0 to 1: lower values retain more source structure, while higher values permit larger changes. Lyrics strength adjusts lyric conditioning, and CFG controls overall text guidance. The uploaded audio supplies the starting latent and temporal extent. Prepare your excerpt before uploading: this tool has no trimming, independent duration control or separate text-to-music mode. Negative conditioning is zeroed internally, with no editable negative prompt. Use seed -1 for random variations or a fixed nonnegative seed to compare settings. Steps, sampler and scheduler control denoising. Advanced latent tonemap multiplier and sampling shift controls change internal guidance and model sampling; retain their source defaults of 1 and 5 for a first run. Sampling shift is not an audio timing offset. MP3 quality offers V0 variable bitrate, 128k and 320k. This controls file encoding rather than the generated composition. For multilingual lyrics, the supplied source guidance calls for manual romanization and language markers such as [zh], [ja] or [ko]; automatic transliteration is absent. The workflow requires ace_step_v1_3.5b.safetensors in the ComfyUI checkpoints model directory. The supplied model source is https://huggingface.co/Comfy-Org/ACE-Step_ComfyUI_repackaged/resolve/main/all_in_one/ace_step_v1_3.5b.safetensors?download=true. The supplied dependency catalog has not resolved this checkpoint.
Kandinsky 5 Lite text to image
Generate an image or image batch from text with the fixed Kandinsky 5 Lite model. Describe the subject, composition, lighting, colors, and style in Positive prompt. Negative prompt can describe unwanted elements or remain blank. No uploaded image is required or supported. For a first run, use the supplied hiker prompt or replace it, keep Width and Height at 1024 pixels and Image count at 1, and leave the advanced settings at 50 steps, prompt guidance 3.5, Euler sampler, Simple scheduler, Schedule shift 3, and Denoise strength 1. Run the tool to receive the generated image batch. The output transport replaces ComfyUI SaveImage and does not permanently save images through that node. Width and Height use increments of 8; adjust them together to set the aspect ratio. Image count controls how many images are generated together. Larger dimensions and batches increase resource use. Advanced Sampling controls change randomness, sampling effort, prompt guidance, and the algorithm and noise schedule used to generate images. Seed -1 selects fresh randomness on each run; use a fixed seed when comparing settings. Schedule shift changes the model's timestep distribution. Denoise strength remains at 1 for the supplied setup because sampling starts from an empty latent. Reducing it does not provide image preservation or an image-to-image mode. The setup uses kandinsky5lite_t2i.safetensors, qwen_2.5_vl_7b_fp8_scaled.safetensors, clip_l.safetensors, and ae.safetensors, plus the Sweet Tea output node pack. These assets must be available in the execution environment. Model and encoder choices are fixed in this tool.
Hoshimachi Suisei 星街すいせい / Hololive
sui1, 1girl, solo, side ponytail, hoshimachi suisei, fingerless gloves, single thighhigh, jewelry, thigh strap, bracelet, buttons, plaid dress, blue choker, blue belt, plaid skirt, mini crown, grey skirt, blue ascot,...
Flux 2 Dev with optional image references
Generate new images from text with Flux 2 Dev, optionally conditioned by one or two reference images. First run: enter a positive prompt and keep the default 1024 × 1024 output, batch size 1, 20 steps and guidance 4. Leave both Use reference stages bypassed; no image uploads are needed in this mode. Run to receive the generated image batch. Seed -1 requests fresh randomization on each run; use a fixed seed to compare settings. For reference conditioning, upload Reference image 1 or Reference image 2 and enable its matching Use reference stage. Either reference works independently, and both can contribute to one generation. Explain in the prompt how the references should inform the result. Each enabled reference is resized to approximately one megapixel before encoding. Advanced reference resolution and interpolation controls affect reference preparation, not output dimensions. References condition a new generation; they are not initial images for denoising. This graph has no negative-prompt or denoise control and does not provide a finished-image upscaling mode. Set output Width and Height in pixels, in multiples of 16. Whenever you change either, also change the corresponding Scheduler width or Scheduler height under advanced Sampling to match. These are separate controls, not automatically linked. Batch size sets the number of generated images. Advanced controls also expose guidance, sampling steps and the sampling algorithm. Requires flux2_dev_fp8mixed.safetensors, mistral_3_small_flux2_fp8.safetensors and flux2-vae.safetensors. Supplied dependency evidence identifies the VAE artifact at https://civitai.com/models/2173342?modelVersionId=2447823 but does not resolve the diffusion-model or text-encoder sources.
Edit video with Runway Aleph 2
Edit existing footage with a text description using Runway Aleph 2. For a first run, upload a source video lasting 2–30 seconds at 30 fps or lower. Replace the example surfing prompt with an English description of what should appear in the result, using 1–1000 characters. Describe the desired changes to subjects, objects, surroundings, lighting, or visual style, then run the tool and inspect the returned video. Leave Seed at -1 to randomize each run, or enter a fixed seed to reuse the generation seed. Public figure moderation offers auto and low settings for recognizable public figures; the template defaults to low. Advanced Video output controls select the output container, codec, and encoding quality. Leave their defaults for a first run. These controls change output encoding, not the requested scene edit. This tool supports text guidance only; the source template's image-guidance and keyframe branches are disconnected from generation. There are no separate output-size, duration, or frame-rate controls. Inspect the result for unintended changes, since the graph does not enforce preservation of untouched details. The source template specifies a logged-in ComfyUI comfy.org API account for the Runway partner node.
TemporalDiff Motion Module
This model was created by CiaraRowles , posted to Civitai.com with permission. Basic Getting Started Tutorial available at the Civitai Education Hub !
Atomix Wan I2V Workflow — Wan 2.2
A Simple Wan I2V workflow
LTX 2.3 video with generated audio
Generate a video with audio from a text description using LTX 2.3. No image, video, or audio upload is required. For a first run, replace the example scene and sound prompt with your subject, setting, action, camera movement, dialogue, and sounds. Review the negative prompt, choose the requested width and height in pixels, and enter a positive duration in whole seconds. Keep the advanced sampling and encoding defaults for the first run. The source defaults request a 900 × 1600 portrait video lasting approximately 20 seconds at 24 FPS. Initial sampling jointly generates video and audio at half the requested spatial size. A fixed 2× latent upscaler feeds a second sampling stage, whose video is decoded for the final result. The returned soundtrack is decoded from the first stage; refinement settings do not replace it. Actual dimensions and duration can differ from the request because of latent spatial and frame-count constraints. Playback FPS controls playback, audio timing, and duration-to-frame conversion. Conditioning FPS is a separate control used by both sampling stages: change both fields together and keep them equal. Use playback FPS within the downstream video's supported 1–120 range. Each sampling stage has its own seed, CFG, sampler, and editable manual noise schedule in advanced settings. Seed -1 randomizes that stage; a fixed nonnegative seed reuses its noise. To reuse both stages' noise, set both seeds. The schedules define sampling intervals directly, so there is no separate steps control. Negative conditioning is connected to both stages, and its effect depends on their CFG settings. The optional fixed distillation LoRA starts disabled, matching the source. Enable its stage to make the LoRA strength operative. The base quantized model, text encoders, VAEs, 2× upscaler, and memory patches are fixed implementation choices. Output settings select the container, codec, bit depth, and compression quality. Start with MP4, automatic codec, 8-bit precision, and automatic CRF (-1). Container and codec must be compatible; 10-bit output also requires compatible encoding and playback. The source's 8 GB hardware claim is not a verified runtime guarantee. Dependency evidence is incomplete: the inventory lists an older upscaler than node 23 selects, and node 26's first text encoder filename has a doubled .safetensors suffix. The graph preserves those source model selections.
Portrait relighting
Generate a relit portrait from a reference image and written lighting directions. For a first run, upload a portrait, keep the default dual rim prompt, leave aspect ratio on auto and resolution at 1K, then run. The prompt requests identity preservation, but this is generative editing: likeness, details and composition can change. Use one editable prompt to explore the five lighting recipes in the source workflow. With a reference, begin with “Maintain the identity of input image 1, modify the scene lighting and shadows to match the following lighting style:” and append one of these recipes: - Dual rim: no front key, twin hard lights from behind left and right, subject silhouetted at center, glowing edge separation on both sides, dramatic dark fill. - Overhead: hard top-down overhead key, harsh light directly above, deep under-eye and jaw shadows, pronounced brow ridge highlight, strong nose shadow pointing straight down, theatrical and sculpted. - Split: hard side key at 90 degrees, one half of the face fully lit, the other half near-black, sharp nose shadow line, minimal fill. - Rembrandt: side key slightly above, deep falloff, subtle rim separation. - Motivated practical: off-axis kicker mimicking an environmental source, strong directional spill, naturalistic shadow logic, atmospheric fill. These are editable prompt recipes within one tool, rather than separate operations or simultaneous outputs. Change the directions freely to request another lighting arrangement. Bypass Source image to generate a new portrait without a reference. In that mode, replace the identity-preservation wording with a description of the subject and desired scene, and choose an explicit aspect ratio. Auto matches the reference aspect ratio when an image is supplied. Resolution offers 1K, 2K and 4K; the node uses Gemini's native upscaler for 2K and 4K. Model selects between the two source-supported Gemini image models. Seed defaults to a fresh random value per run; a fixed value requests best-effort consistency, not deterministic reproduction. Response mode controls whether the provider also generates text, but this Pipe delivers the image only because the source text output is unconnected. Generation instructions provide advanced behavioral guidance; retain instructions to produce an image when editing them.
Topaz creative upscale
Create a creatively enhanced, resized image with Topaz Reimagine. Upload a source image and choose Scale factor. For a first run, use the default 4× scale, leave Prompt empty and keep Creativity at 2. A 4× scale requests four times the source width and height, approximately 16 times as many pixels; a scale of 1 requests the original dimensions. Add an optional prompt to guide the desired details or appearance. Creativity accepts values from 1 to 9. Preserve colors enables preservation of the original palette. Subject detection offers All, Foreground and Background modes. For portraits, enable Enhance faces to use Face creativity and Face sharpness, both ranging from 0 to 1. Face sharpness controls how sharp enhanced faces are relative to the background. Preserve facial identity is a separate option and defaults to off. Output dimensions follow the source aspect ratio, subject to pixel rounding; independent width and height choices are unavailable. The size calculation branch supplies dimensions, while the original image goes directly to Topaz. Crop to fill remains off because this workflow has no independent aspect-ratio choice. Uploaded transparency is not preserved: only the loader's image output is connected, without its alpha mask.
Wan 2.1 video inpainting
Generate an edited video segment with SAM3 object selection and Wan 2.1 VACE masked video conditioning. For a first run, upload a source video, replace Object to edit with a description of existing content in that clip, and replace Edit prompt with the desired replacement, scene and motion. Leave Reference image disabled for text-only editing. To add appearance guidance, enable that stage and upload an image. Start frame is a zero-based source-frame index. Frame count requests the segment to edit, and the actual selected frame batch determines generation length. Begin with Start frame 0 and the default 81 frames when the clip has enough frames. Approximate segment duration is frame count divided by the source frame rate. This tool outputs the generated segment, without reinserting it into the full source video. Output width and height are in pixels and are rounded down to multiples of 16. Use positive dimensions of at least 16. Center framing crops while preserving proportions; Disabled stretches to fit. SAM3 selects content after resizing. Mask margin expands that selection in resized-frame pixels; negative margins shrink it. Detection threshold and refinement passes adjust selection. Combined masks are the normal video setting; separate object masks have no dedicated alignment stage here. Fast sampling selects the fixed CausVid LoRA and the Fast steps and Fast CFG controls. Turning it off selects the base model and Standard steps and Standard CFG. CausVid LoRA strength affects only fast mode. Both modes share the seed, sampler, scheduler, denoise strength and sampling shift. Use seed -1 for randomization. VACE control strength is separate from prompt guidance. Keep batch size 1 for the normal single-sequence output; larger batches are not split into separate video outputs. The output uses the source video's frame rate and passes through its audio. Audio is not explicitly trimmed or offset to match Start frame and Frame count, so synchronization and duration may differ for a selected segment. Generated frames are decoded directly, without a final composite that guarantees unchanged pixels outside the mask. Video encoding controls select the output container, codec and compression setting. The workflow uses fixed SAM3, Wan VACE, text-encoder and VAE models; fast mode additionally uses the fixed CausVid LoRA. These model assets must be available to the execution environment. The supplied dependency evidence marks them missing; this review establishes the connected tool and its guidance, not a completed runtime test.
Wan 2.2 text to video
Generate one silent video from a text description with Wan 2.2. No uploaded image or video is required. For a first run, edit the positive prompt to describe your subject, movement, camera, lighting and setting, then keep the 640 × 640 size, 5-second duration and 16 FPS defaults. The negative prompt independently discourages unwanted content and visual traits; its supplied Chinese text includes artifacts, static frames and clutter. Standard mode uses 20 total steps, CFG 3.5 and a handoff at step 10. Enable Lightning mode to select both Lightx2v LoRAs together with the separate Lightning settings: 4 total steps, CFG 1 and a handoff at step 2. Only the selected mode's steps, CFG and handoff settings apply. LoRA strength controls apply only in Lightning mode. Each handoff must remain strictly between zero and its mode's total step count. The downstream samplers accept 1–10000 total steps and CFG values from 0–100. Width and height are pixels in multiples of 16. Use a positive duration and a frame rate from 1–120 FPS. Requested frame count is floor(duration × FPS) + 1; choose a combination whose floored product is a multiple of 4 for the expected 4n + 1 frame layout. FPS also sets playback speed. The defaults request 81 frames, giving approximately 5.06 seconds of encoded playback. Larger dimensions, longer duration and higher FPS increase generation work and memory use. Seed -1 randomizes each run; use a fixed seed to compare settings. Advanced initial and finishing sampler, scheduler and schedule-shift controls apply in both modes. The initial stage adds noise and runs the high-noise model; the finishing stage continues the same latent without adding new noise. Both stages use the selected total step count, with the handoff dividing their work. The supplied sampler, scheduler and shift values match across stages. Keep file type, codec and encoding quality on automatic selection for a first run. Explicit container and codec choices require compatible encoding and playback support. Video assembly defaults to 8-bit; optional 10-bit assembly can retain smoother gradients but requires downstream support. CRF controls encoding quality separately from generation. Latent batch size remains one; this tool returns one video and does not offer separate batch files. Audio is not generated.
