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
12 shown of 34,764
Wan 2.1 edge-guided video
Generate a silent video guided by source-video edges, a start image, and a text prompt using Wan 2.1 Fun Control. For a first run, upload a source video and a start image, then replace the positive prompt example with your desired subject, action, and style. The source video supplies frame-by-frame Canny contours. The start image supplies both initial-frame and visual conditioning. The negative prompt lists unwanted appearance or motion; its packaged default is in Chinese. Keep 960 × 544 pixels, 77 frames, 16 FPS, batch size 1, full denoise, and the packaged sampling settings initially. This gives approximately 4.8 seconds of video. Dimensions use multiples of 16 and frame counts follow 1 + 4n. Choose a source clip with enough frames for the requested sequence. There are no source trimming or frame-rate conversion controls. Low and high edge thresholds determine which contours guide generation. Lower thresholds include weaker edges; higher thresholds favor stronger contours. Keep the low threshold below the high threshold. Reference crop changes framing for visual conditioning only, independently of the initial-frame input. Frames determines the generated sequence length. FPS determines playback speed and duration without generating additional frames. Source FPS and source audio are not connected to output playback. The only selected output is the generated video; no separate edge preview is returned. Leave file type and codec on auto for a first run. Manual choices must be compatible with each other and your playback software. The optional 10-bit setting also requires compatible encoding and playback. Advanced sampling controls include seed, steps, prompt guidance, sampler, scheduler, denoise, and schedule shift. Seed -1 requests fresh random noise; a fixed seed lets you compare settings with the same starting noise. Keep denoise 1 and batch size 1 for normal use; batch variants have no separate video output branches. Skip-layer guidance exposes independent double-block and single-block layer lists, strength, rescaling, and a sampling-progress interval expressed as fractions from 0 to 1. Keep its start below its end. Attention adjustments independently scale structural and temporal self-attention and cross-attention. Either adjustment stage can be bypassed while preserving the model path; its controls apply only while enabled. The model, text encoder, vision encoder, and VAE remain fixed to the packaged Wan 2.1 Fun Control setup.
Gemini two-image editing
Generate an edited image from a source scene, a second visual reference and written instructions using Gemini. Upload both required images: the scene to change as Source image and an object, accessory, clothing item or other visual guide as Reference image. The source image is sent first and the reference second, together to Gemini. Explain each image's role, what to change and what to preserve. For a first run, upload a scene containing a person and a separate accessory reference. Replace the example Edit instructions with a request such as: “Use the first image as the scene. Put the accessory from the second image on the person, keeping the surrounding scene unchanged.” Choose an aspect ratio and resolution, then run the Pipe. Aspect ratio controls output proportions; auto asks Gemini to follow the input image's aspect ratio. Resolution offers 1K, 2K and 4K targets; 2K and 4K use Gemini's native upscaler. Model offers Gemini 3 Pro Image Preview or Nano Banana 2 (Gemini 3.1 Flash Image). Seed -1 randomizes each run; a fixed seed requests best-effort repeatability without guaranteeing identical images. System instructions set foundational provider behavior alongside the edit prompt. The default emphasizes producing an image. Provider response defaults to IMAGE; IMAGE+TEXT additionally requests text, but only the generated image is returned by this Pipe. Preservation of faces, objects and unchanged regions is a prompt request, not a guarantee. Both loaders send their IMAGE outputs; uploaded alpha masks are not connected, so uploaded transparency is not preserved through this path.
ComfyUI-ComfyCouple
Attention Couple made easier for ComfyUI.
Kling reference-guided video transformation
Transform a source video using a reference image and written editing instructions, returning one generated video. Upload both a source video and a reference image. Both uploads are required: this workflow always supplies the image to Kling and has no video-only mode. For a first run, choose a video showing the subject you want to change and an image showing the desired character or appearance. Replace or adapt the example prompt, which asks for a dancing person to become a rabbit while retaining the movement, background and camera sequence. Describe the changes, how to use the reference image, and what should remain consistent. Include unwanted content in the same prompt; there is no separate negative prompt. Choose landscape 16:9, portrait 9:16 or square 1:1, a generated duration of 3–10 seconds, and 720p or 1080p resolution. Keep original sound requests retention of the source audio. Model selection offers kling-video-o1 and kling-v3-omni. Motion, likeness, scene consistency and sound retention are requests rather than guarantees. Leave Seed at -1 to randomize it on each run. The provider uses the seed to control re-running; even a fixed seed does not make the generated video reproducible. Advanced video output settings offer file type and codec selection; both default to auto. Generation sends the supplied media and prompt to Kling and uses a paid provider call.
Text to video with SDXL and SVD
Generate a starting image from text with SDXL, then animate it into a short, silent video with SVD XT. For a first run, keep the supplied settings and example prompts, or replace the positive prompt with your subject, setting and desired appearance. Use the negative prompt for unwanted image features. Run the Pipe to receive the video. No image upload is required or exposed: the starting image is generated within this workflow. The selected output is video, without an independent still-image output or audio. The default video is 1024 × 576 pixels with 25 frames at 10 playback FPS, giving a 2.5-second clip. Change frame count and playback FPS together to control duration. Conditioning FPS defaults to 6 and influences SVD generation independently of playback FPS. Motion bucket and starting image noise are additional SVD conditioning controls; they do not specify an exact motion trajectory. Positive and negative prompts affect only the SDXL image stage. SVD animates that generated image and does not receive text conditioning directly. Starting image dimensions and video dimensions are independent; both default to 1024 × 576. Starting image dimensions are available in advanced settings. Image and video sampling have independent seeds, steps, guidance, samplers, schedulers and denoising amounts. Seed -1 randomizes that stage. Fix both seeds when comparing runs with the same starting image and animation noise. Video guidance varies linearly between the minimum guidance setting and the video sampler's maximum CFG; keep minimum at or below maximum for that relationship. Retain the default denoising amounts of 1 for the first run. Advanced playback settings include 8-bit or 10-bit video. The source notes possible player and downstream compatibility limitations for 10-bit. Video output settings offer automatic selection or explicit container and codec choices, plus a separate encoding CRF setting. These encoding controls do not change the generation prompts or sampling guidance. The workflow uses sd_xl_base_1.0.safetensors and svd_xt.safetensors from the source's Stability AI model references, loaded from the checkpoints directory, plus Sweet Tea's video output node. Those models must be available to the runtime. The selected graph always includes both image generation and video generation; it has no upload mode or bypassable stages.
Depth-guided image generation with Flux
Generate a new Flux image from a prompt using the estimated depth and spatial layout of an uploaded source image. For a first run, upload a source image, replace the example prompt with a description of your desired subject and appearance, and leave the source defaults in place: 1 megapixel, depth LoRA strength 1, prompt guidance 10, and 20 generation steps. Seed -1 randomizes generation; use a fixed nonnegative seed when comparing settings. The selected output is the final generated image, not a separate depth-map result. The source is resized while retaining its aspect ratio, then Lotus estimates its depth. The inverted depth map conditions Flux.1 dev with its fixed depth LoRA. Output dimensions follow the resized source, subject to latent encoding. Changing output megapixels or resize interpolation also changes the input to depth estimation. Advanced controls separate depth LoRA strength, Flux prompt guidance and final generation sampling. Generation denoise acts on the conditioning-created latent, so it does not directly preserve source colors or texture. Sampler CFG is independent of Flux prompt guidance. Negative conditioning is zeroed and there is no separate negative prompt. Depth estimation has its own sampler, steps, scheduler and denoise controls. Its source configuration uses Euler, one step, normal scheduling, full denoise and disabled added noise; a downstream stage fixes the first sigma at 10000. These settings affect the estimated depth used by generation. Depth estimation and the depth LoRA remain integral stages; there is no bypass or alternative generation mode. The workflow uses the configured Flux.1 dev, Flux depth LoRA, CLIP-L/T5 encoders, Flux VAE, Lotus depth model and Lotus VAE. Source RGB supplies depth guidance rather than direct appearance copying, and uploaded alpha transparency is not carried through to the result.
Kling reference-guided video edit
Create an edited video from a source clip, written instructions and one reference image. Upload both required media inputs and replace the example edit instructions with a description suited to your uploads. State what to change, what to retain and what to avoid. Use [@Video] to refer to the submitted clip and [@Image] to refer to the reference image. For a first run, leave Start time and Duration at 0 to submit the whole video, keep kling-video-o1 and 1080p, and choose whether to retain original sound. Under Clip selection, enter Start time and Duration in seconds to edit a shorter segment. Duration 0 submits the remainder of the video. Require full duration raises an error when the requested duration is unavailable. The edited video follows the submitted clip's length; Duration selects source footage rather than a new generation length. Model also offers kling-v3-omni, and Resolution offers 720p. This graph supplies one required reference image; it does not expose the service's additional reference slots or a reference-free mode. Requests to preserve lighting, background or movement guide the edit without guaranteeing unchanged pixels or motion. Seed -1 randomizes each run; a fixed seed does not make this provider's results reproducible. Video output returns the edited clip for preview. Its File type and Codec choices control final encoding, with automatic file type and H.264 as defaults. Editing sends the submitted media to the paid Kling service.
LTX-2.3: Text to video with audio
Create one video with generated audio from a scene prompt using LTX-2.3. Describe the scene, actions over time, camera movement, visual details, and desired sounds or dialogue. This graph accepts no image, video or audio uploads. For a first run, replace the example scene prompt and retain the defaults: requested size 1280 × 720, duration 5 seconds, 25 FPS, Enhance prompt enabled, Text-only mode enabled, and automatic output format and codec. The workflow generates video and audio together at half the requested spatial dimensions, applies fixed 2× video latent upscaling, then refines both and returns the combined result. Refinement runs every time. Enhance prompt chooses a generated rewrite as the video conditioning text. Disable it to encode your original wording directly. Advanced enhancement controls include sampled or unsampled text generation, token budget, sampling filters, template selection and text encoder LoRA strength. Temperature and the enhancement seed are relevant to sampled rewriting. Thinking mode is available only if the loaded text model supports it. Width and height request the final spatial size; latent alignment may affect realized dimensions. The initial latent receives half each dimension and has a source minimum of 64 pixels with a 32-pixel step. For aligned custom sizes, use final dimensions of at least 128 pixels in multiples of 64. Swap width and height for portrait framing. The retained 1280 × 720 default is preserved; arbitrary dimensions are not guaranteed to produce an exact requested size. Duration is positive whole seconds. Requested frame count is duration × FPS + 1. Use 1–120 FPS and keep the requested frame count at or below the audio latent's source limit of 1000. Video latent length has an eight-frame step, so frame alignment can affect the realized duration; for aligned custom timing, make duration × FPS a multiple of eight. Frame rate also controls video conditioning, generated audio timing and final playback. Text-only mode bypasses solid-color first-frame conditioning in both stages. Turn it off to use the chosen solid color, independently adjustable initial and refinement conditioning strengths, and conditioning compression. Color is a packed RGB integer: 0 is black and 16777215 is white. This mode uses a generated color image rather than an uploaded reference. Initial generation and refinement have independent seeds, CFG values, samplers and comma-separated noise schedules. The distilled LoRA strength affects both stages. Keep the supplied LoRA strength, Euler samplers and schedules for the first run. Negative conditioning is supplied to both stages, but ordinary classifier-free negative guidance has no effect at their default CFG of 1; increasing a stage's CFG makes its negative conditioning relevant. Generation, refinement and sampled prompt enhancement have separate seeds. Use -1 for fresh randomization, or fix every relevant seed when comparing settings. Advanced output controls select the video container, codec and encoding CRF. Keep their automatic settings and default CRF for the first run; the supplied evidence does not establish compatibility for every explicit container and codec pairing.
MiniMax H3 video with audio
Generate a widescreen video with accompanying audio from a text description using MiniMax H3. Describe the scene, action, visual style, camera movement, dialogue, sound effects, and music in the Scene and audio prompt. There is no image upload or separate negative prompt in this Pipe. For a first run, replace the example with a short scene suitable for the two-second duration default and retain the sampling defaults: 20 steps, res_multistep, simple scheduler, and denoise 1. The saved example describes five seconds; shorten its storyboard or increase Duration before using it. Prompt timestamps and camera directions describe your intent rather than guaranteeing exact cuts or timing. Duration is requested in seconds. At fixed 24 fps, the workflow rounds seconds × 24, applies a minimum of five frames, then snaps upward to a valid 17k+5 frame count. Two seconds produces 56 frames, about 2.33 seconds, so the delivered clip can be longer than requested. Output size is not adjustable here: the retained source resolution selector uses its saved 16:9, 0.4-megapixel settings. Advanced controls change generation: Seed -1 chooses fresh random starting noise each run; a fixed seed lets you compare changes with the same starting noise. Steps sets the sampling step count, Sampler selects the sampling method, Scheduler distributes noise levels across those steps, and Denoise selects the fraction of the noise schedule used. Lower denoise does not preserve an uploaded image because this Pipe has none. Video frames and audio are decoded from the same sampled result and delivered together as one video at 24 fps. Use the default 8-bit output for compatibility. The supported 10-bit option can retain smoother gradients but requires compatible encoding and playback. The configured MiniMax H3 diffusion model, Qwen3VL text encoder, video VAE, and audio VAE must be available in the execution environment. This Pipe does not expose model selection, image conditioning, end frames, or an audio-only output.
Stable Audio 3 audio creator
Create an MP3 audio clip from a written description using Stable Audio 3 Medium. No uploaded audio is required. Describe the sound source, instruments or materials, mood, rhythm, and production character you want. For a first run, keep the tropical-house example, Music category, Expand prompt enabled, and Duration at 150 seconds. Keep the source sampling settings: 8 steps, CFG 1, LCM sampler, simple scheduler, and Denoise 1. Run the tool and listen to the MP3 result. MP3 quality changes encoding and file size rather than the generated performance. Prompt expansion uses Qwen to elaborate your description. Category must be exactly Music, Instrument, SFX, or One-shot. Music requests an instrumental track; Instrument requests instrument-focused material; SFX describes effects or ambience; One-shot requests a short isolated sample. These are prompting modes, with no stem extraction or guaranteed seamless looping. Category does not change Duration automatically: set a shorter duration yourself for short effects or samples. Turn Expand prompt off to send your original description directly to audio conditioning. Category and expansion tuning then do not affect the selected conditioning. This selection does not promise that the runtime skips loading or executing every expansion dependency. Duration controls the empty audio latent in seconds. The connected latent node supports 1–1000 seconds. Use whole seconds because the expansion request converts Duration to an integer. Category templates contain their own length suggestions, so expanded wording can conflict with your requested duration; the latent duration remains controlled by Duration. Disable expansion when you want direct control over prompt wording. Text requests do not guarantee exact BPM or event timing. Audio seed and Prompt seed control separate random stages. Use -1 for fresh randomization. To compare audio settings with consistent wording, disable expansion or fix Prompt seed and retain the same expansion settings. Prompt sampling off disables text sampling; temperature and token-sampling controls apply to sampled expansion. Expansion length limit is measured in text tokens, not audio seconds. Thinking mode depends on model support, and generated text has no separate reasoning-text cleanup stage before audio conditioning. Negative prompt describes unwanted sounds. At the default CFG of 1, it does not contribute a positive-versus-negative guidance difference. Higher CFG changes that guidance behavior. Retain Denoise 1 for the first run because generation starts from an empty latent. Audio batch controls the number of latents generated together; larger batches require more memory. The runtime must provide the fixed source models: stable_audio_3_medium.safetensors in checkpoints and t5gemma_b_b_ul2.safetensors in text_encoders, supplied by https://huggingface.co/Comfy-Org/stable-audio-3; and qwen3.5_2b_bf16.safetensors in text_encoders, supplied by https://huggingface.co/Comfy-Org/Qwen3.5. These source dependencies are retained even though the supplied catalog has not resolved them.
Generate a background-removal mask
Generate a segmentation mask from an uploaded image with BiRefNet. The mask is returned as a separate output for compositing or downstream image processing.
Wan 2.2 image to video
Generate one silent video from a starting image and a motion prompt using Wan 2.2's paired 14B image-to-video models. For a first run, upload an image, replace the example prompt with the subject's action and camera movement, and keep 640 × 640, 5 seconds, 16 FPS, batch size 1, and Fast mode off. Choose width and height in multiples of 16 to suit the image's composition. An upload is needed while the Source image stage is enabled. Disable that stage to generate without starting-image conditioning, and describe the subject and scene in the prompt as well as their motion. This mode retains the same image-to-video model pair. Uploaded transparency is not preserved by the connected RGB image output. Fast mode selects both Lightx2v acceleration LoRAs and the fast sampling controls. Defaults change from 20 steps, CFG 3.5, and handoff step 10 to 4 steps, CFG 1, and handoff step 2. Only the selected mode's steps, guidance, and handoff apply. Keep total steps between 1 and 10000, guidance between 0 and 100, and each handoff greater than zero and below its total steps. The two LoRA strengths apply only in fast mode. Both stages have independent sampler, scheduler, and noise-schedule shift controls; the supplied recipe uses matching Euler samplers, simple schedulers, and shifts of 5. Duration controls generation length. The graph requests floor(duration × FPS + 1) frames and uses the same FPS for playback. Use a positive duration and 1–120 FPS; choose duration × FPS as a multiple of four for the model's 4n + 1 frame convention. Defaults request 81 frames at 16 FPS, approximately five seconds. Increasing resolution, duration, or FPS increases generation work. Keep batch size 1 for one clip; larger batches still pass through a single video-assembly branch rather than separate clip outputs. Use seed -1 for a new random seed each run or a fixed seed to compare changes with the same initial noise. Advanced controls include a negative prompt for unwanted traits; its influence depends on guidance. The tool returns one video through Sweet Tea, using automatic container and codec selection by default. Available containers are mp4, mkv, and webm; explicit codecs are h264 and av1. Keep automatic selection unless you need a particular compatible combination. Eight-bit output is the default; 10-bit output requires compatible encoding and playback. The graph has no audio input, end-frame conditioning, or upscaling stage.
