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 42.580
Video pose maps and overlays
Upload a video to produce a rendered skeleton map or pose overlay with person detection boxes. The output uses the source frame rate and available audio. Analysis frames and background frames share the same longer-edge resize, preserving aspect ratio. For a first run, upload your source video and keep the 1024-pixel longer edge, maximum of two people, normal blend mode, and pose blend of 0.5. Run the tool to obtain the visualization video. In normal mode, pose blend 0 shows resized source frames with detection boxes, 0.5 mixes those frames with the pose map, and 1 shows the standalone pose map. Detection and pose extraction still run at every blend setting. Multiply, screen, overlay, soft light, and difference are optional layer-combination modes; the normal-mode endpoint examples do not apply to them. Maximum people limits detections in each frame by descending confidence. It does not select or track a particular identity. Detection confidence filters person boxes used for both the overlay and pose extraction. Keypoint confidence filters the points drawn in the pose map. Lower thresholds can admit uncertain detections or points. Under Pose drawing, choose body, hands, face, head, and feet independently, and adjust skeleton line width and face point size in pixels. These controls change the pose layer; their visible contribution depends on the blend settings. The output is a rendered video, not a keypoint data export. Set the longer edge to a positive integer number of pixels. Larger dimensions increase processing demand. Processing batch size sets how many frames pose extraction processes together; reduce it if extraction exhausts memory. For a first run, retain automatic file type and codec with 8-bit output. Optional MP4, MKV, or WebM file types and H.264 or AV1 codecs must be mutually compatible and supported by your player. Ten-bit output can represent smoother gradients but requires compatible encoding and playback.
Temperature, tint, vibrance and saturation
Create a color-adjusted image with temperature, tint, vibrance and saturation controls. For a first run, upload a source image and run with all four saved controls at 0 to preview the image without color adjustments. Change a control and run again to compare the result. All controls use unitless -100 to 100 scales; temperature is not measured in Kelvin. Positive temperature adds red and reduces blue, warming the image; negative temperature cools it. Positive tint adds green while reducing red and blue; negative tint shifts toward magenta. These adjustments combine and are clamped before vibrance is applied. Positive vibrance boosts less saturated colors more, with some reduction for warm colors. This is a color heuristic rather than skin detection or guaranteed skin-tone preservation. Negative vibrance reduces color intensity, reaching grayscale at -100. Saturation then adjusts overall color intensity: positive increases it, negative reduces it, and -100 produces grayscale. Increasing saturation cannot restore colors already removed by vibrance at -100. The four controls combine in one output, in the order temperature and tint, vibrance, then saturation. Strong adjustments can clip colors and lose detail. There are no additional selectable modes. Output dimensions follow the source image. Uploaded transparency is not preserved because the image loader's separate alpha mask is unused. This output applies only these four adjustments; other effects in this family are separate tools.
bad-picture negative embedding for ChilloutMix
March 17, 2023 edit: quick note on how to use a negative embeddings. You download the file and put it into your embeddings folder. Since I use A1111 Webui, that is stable-diffusion-webui/embeddings. If you use a...
SD3.5 depth-guided image
Generate a prompted SD3.5 Large image guided by depth estimated from a source image. Provide a source image and describe the image to generate. Output size retains the source aspect ratio at the selected megapixel budget. Depth guidance controls its strength and the portion of generation during which it applies. Advanced controls tune generation, randomization and Lotus depth estimation. The inverted depth map supplies both ControlNet guidance and the initial generation latent; lowering Denoise retains more of that depth-map initialization. Negative conditioning is zeroed automatically, so there is no separate negative prompt. The output is the final generated image.
Ideogram 4 image generator
Generate images from scene captions using the fixed Ideogram 4 conditional and unconditional model pair. No uploaded media is required. For a first run, keep the supplied skateboard poster caption, 9:16 portrait aspect ratio, 1 megapixel resolution, Default sampling preset, and image count of 1, then run the Pipe. Replace Positive prompt with your own scene afterward. Prefer a structured JSON caption like the example, describing the scene summary, style, background, and individual objects. Plain natural language is also supported. Optional caption bounding boxes use normalized [y1, x1, y2, x2] coordinates from 0 to 1000. These are prompt content, not separate placement controls. There is no separate negative prompt: the unconditional branch receives zeroed conditioning. Prompt refinement is not automatic; paste any externally prepared caption into Positive prompt. Aspect ratio and megapixels jointly determine the output dimensions. Each dimension is rounded up to a multiple of 16 with a minimum of 256 pixels, so the resulting ratio and pixel count can differ from the request. Sampling presets coordinate the step count and scheduler settings: Quality uses 48 steps, mu 0, and std 1.5; Default uses 20 steps, mu 0, and std 1.75; Turbo uses 12 steps, mu 0.5, and std 1.75. Advanced controls expose image count, seed, sampler, and guidance. Seed -1 chooses fresh random noise for each run; use a fixed seed to reuse the initial noise while comparing settings. Euler is the source sampler default. Main prompt guidance defaults to 7, with override guidance of 3 during the final 30% of sampling. Override start and end are fractions from 0 to 1; keep start below end. Increasing resolution, image count, or sampling steps increases resource use. The workflow requires ideogram4_fp8_scaled.safetensors and ideogram4_unconditional_fp8_scaled.safetensors in diffusion_models, qwen3vl_8b_fp8_scaled.safetensors in text_encoders, and flux2-vae.safetensors in vae. Dependency evidence currently records these files as missing. Their supplied installation sources are the Comfy-Org Ideogram-4, Qwen3-VL, and flux2-dev repositories on Hugging Face. These model choices are fixed in this Pipe. The disconnected caption-construction template does not invoke an LLM or feed image generation. Describing transparency in a caption does not establish an alpha-channel output capability.
Product scene compositing
Generate a new product-scene composite from a product photo and a background scene using Seedream 4.5. Both uploads are required: the scene becomes reference 1 and the resized, padded product becomes reference 2. For your first run, upload both images, replace the example Product description with your product's shape, colors, branding, materials, texture, and reflectivity, then edit Compositing instructions. Replace its bracketed material examples and specify the placement, scale, and lighting you want. Keep {product_description} in the instructions to insert your separate description. Keep the default 2048 × 2048 output and Related-image mode disabled to generate one composite. Output size controls generated dimensions. Width and Height apply only when Output size is Custom; both are measured in pixels. Product reference preparation resizes and pads the product before generation. Reference padding changes that input preparation, not the generated scene background; Resize interpolation changes how the product reference is resampled. Set Related-image mode to auto and describe the desired group in Compositing instructions to request related images. Maximum images is an upper bound, not a guaranteed count. Use at most 13 generated images because the two references count toward the provider's 15-image total. Require complete results aborts when requested images are missing or return an error; disable it to return available images from a partially successful request. Seed -1 randomizes each run; a fixed seed reuses the generation's randomization setting. AI watermark requests an AI generated watermark on the output. This is generative compositing: preservation of logos, colors, texture, and shape is requested through instructions rather than guaranteed. Review product and brand details in the result. Uploaded alpha masks are not connected, so transparency is not used as a compositing mask. Running this Pipe requires access to the Seedream provider through ComfyUI.
Wan 2.7 text to video
Generate one 2–15 second video from a text prompt using the Wan 2.7 partner API. Configured partner API access is required; generation sends your prompt to the provider and uses provider billing. For a first run, keep the supplied florist prompt, 720P resolution, 16:9 aspect ratio and five-second duration, then run the tool. Replace the prompt to describe your scene, subject movement, lighting, visual style and camera movement in English or Chinese. Optionally enter a negative prompt describing unwanted elements; it can remain empty. Choose 720P or 1080P resolution. Aspect ratios 16:9 and 4:3 are landscape, 9:16 and 3:4 are portrait, and 1:1 is square. Duration accepts whole seconds from 2 to 15. AI prompt enhancement is enabled by default and allows AI assistance to expand your wording. Disable it to submit your original wording without enhancement. Seed -1 randomizes each run; a fixed seed supplies a generation seed without guaranteeing identical provider results. Enable Watermark to request an AI-generated watermark. This tool has no image or audio upload. With no audio supplied, the model automatically generates matching background music or sound effects. The generated video is returned to Sweet Tea with automatic container and codec selection by default.
Mage-Flow text to image
Generate square images from a text description using the configured Mage-Flow INT8 model. No uploaded image is required or supported. For a first run, replace the example positive prompt with your subject, composition, lighting and style. Leave the negative prompt empty, choose one image, and keep the supplied sampling defaults: 30 steps, prompt guidance (CFG) 5, Euler sampler, Simple scheduler and denoising strength 1. The source resolution stage is configured for square output at approximately 1024 × 1024 pixels; this Pipe has no output-size control. Use the negative prompt to describe unwanted content or visual traits. Images per run controls the batch size; larger batches require more memory. The seed defaults to -1 for a new random seed each run. Use a fixed seed to revisit a variation with otherwise unchanged settings. Advanced sampling exposes steps, prompt guidance, sampler, scheduler and denoising strength. Steps set the number of denoising iterations. CFG controls adherence to the prompt; excessive guidance can degrade the image. The sampler chooses the denoising algorithm, and the scheduler controls the progression of noise levels. Denoising strength ranges from 0 to 1 and defaults to full denoising; reducing it does not enable image editing or preserve an uploaded image. The model is fixed. Other model variants, image-editing modes and reference-image inputs are not selectable in this Pipe. Generated images are returned as a batch.
[LamaPanama] Waven Chibi Style
Hypernetwork for generating images stylized as waven chibi characters. Actually produces good results even in different proportions of strength. Trained with danbooru tags.
DALL·E 2 image creation and editing
Create one to eight square images with DALL·E 2. Write a prompt describing the desired result, choose an output size, and set the image count. Source image is enabled by default. Upload an image to edit; its RGB image and alpha-derived mask both feed the provider. Transparent regions supply the replacement mask, whose white areas are edited. This interface has no separate mask upload and does not promise transparent output. For a first text-only run, disable Source image, replace the example prompt, and generate one image at 1024×1024. Disabling Source image removes both image and mask conditioning while leaving generation active. The workflow uses a partner API node and requires an authenticated runtime with sufficient API credits. Seed is retained as a hidden runtime field because the provider documents it as unimplemented; it cannot reliably reproduce results.
Wan 2.1 image-to-video with optional end frame
Generate a silent video from a starting image and a text prompt using Wan 2.1 Fun InP 1.3B, optionally guiding the ending with a second image. For a first run, upload a start image and replace the example positive prompt with a description of your subject, scene, movement, and camera behavior. End frame is enabled by default: either upload an end image or bypass End frame to use the start image alone. The start image remains required in both modes because it also supplies visual conditioning. Images guide generation; an exact transition or exact endpoint match is not guaranteed. Begin with the supplied 480 × 768 dimensions, 81 frames, 16 playback FPS, and 20 sampling steps. This gives approximately 5.1 seconds of video. Width and height are pixels in multiples of 16; frame counts follow 4n + 1, such as 49, 81, or 121. Choose an input aspect ratio close to the output dimensions. Larger dimensions and more frames increase processing and memory requirements. FPS sets playback speed without adding or interpolating frames; duration is frame count divided by FPS. Keep Batch size at 1 initially: larger batches do not have separate clip output paths. The negative prompt describes unwanted appearance and motion; the supplied Chinese text discourages artifacts, poor anatomy, static imagery, and clutter. In Sampling, seed -1 chooses fresh random noise for each run; use a fixed seed to compare settings. Keep the supplied sampler, scheduler, guidance scale, and denoise fraction initially. Denoise is not a direct start-image preservation strength. Reference crop affects visual conditioning from the start image, independently of output dimensions. Skip-layer guidance is enabled by default and can be bypassed for comparison; its strength, start/end fractions, rescaling, and layer selections apply only when enabled. Noise schedule shift and the independent attention multipliers expose additional model tuning. Their effects are separate from playback timing; retain the packaged settings for the first run. Video playback also offers 8-bit or 10-bit depth. The supplied 8-bit setting favors compatibility; 10-bit supports smoother gradients but may be unsupported by some players or downstream tools. The output controls offer automatic encoding or an explicit file type (MP4, MKV, WebM) and codec (H.264, AV1). Automatic format and codec are the defaults. No audio input is connected, so the result is silent.
SDXL text to image with refinement
Generate one image or a batch from text using SDXL 1.0 Base followed by SDXL 1.0 Refiner. No source image is required. The fixed checkpoints sd_xl_base_1.0.safetensors and sd_xl_refiner_1.0.safetensors must be available to the runtime. For a first run, describe your scene in Positive prompt and copy that description into Refiner positive prompt. Enter unwanted features in both negative prompts, then run with the default 1024 × 1024 dimensions and Image count of 1. All four prompt fields are independent: editing a base prompt does not update its refiner counterpart. Keep their intent aligned, or deliberately change the refiner wording to emphasize finishing details. Width and Height are in pixels and use multiples of 8. Image count generates a batch using the same prompts. Larger dimensions and batches require more memory. In advanced settings, Seed controls the initial base noise: -1 selects a new seed each run, while a fixed seed lets you compare changes using the same initial noise. Guidance and sampler choices are independent for Base and Refiner. The default schedule has 25 total steps: Base performs the first 20, passes its partially denoised latent onward, and Refiner completes the remaining five without adding fresh noise. Refinement is integral to this output path and cannot be bypassed through this interface. Advanced schedule controls must be adjusted together. Keep Base schedule steps equal to Refiner schedule steps, use matching schedulers, and set Base handoff step equal to Refiner handoff step below the total schedule length. The schedule-step fields describe the full schedule for each stage, not how many steps that stage executes. Begin with the coordinated defaults before changing the schedule.
ACE Step 1.5 music generator
Describe the genre, instrumentation, vocals, and arrangement, then provide lyrics and musical settings. The saved example is a two-minute K-pop song. Output duration sets the audio length; conditioning duration separately tells the music encoder the intended length, so keep both values equal for ordinary use. Music seed controls conditioning variation, while Seed controls diffusion noise. Negative conditioning is automatically zeroed and has no separate negative prompt. The fixed Turbo model uses eight sampling steps by default and exports MP3 audio.
SAM3 object segmentation
Find objects by name in an uploaded image and return detection masks plus a companion image with detection-derived transparency. For a first run, upload an image, enter a short object description such as "guitar", and keep the detection threshold at 0.5, refinement passes at 2, and separate object masks disabled. The prompt selects objects in the source image; it does not generate new image content. Use a compatible SAM3 checkpoint installed in the runtime; the default is sam3.1_multiplex_fp16.safetensors. Keep prompts within the model's 32-token limit. Separate subjects with commas; append :N to limit detections for a subject, for example "eye:2, window panels:4". Raise the detection threshold to reject weaker matches, or lower it to include weaker matches. Refinement passes range from 0 to 5; zero uses raw detector masks, while additional passes apply SAM decoder refinement. By default, detections share one combined mask. Enable separate object masks for per-object results; output count then depends on detections. The companion image joins the original RGB image with the detection masks using ComfyUI's mask-to-alpha convention: detected regions become transparent. Uploaded transparency is unused, so this output does not preserve the source alpha or provide an isolated opaque object cutout. In the advanced Model section, choose a compatible SAM3 checkpoint from the runtime checkpoint selector. Other installed checkpoints may appear, but ordinary image-generation models are unsuitable. This graph supports text prompting; point prompts and bounding-box inputs are not connected.
Sonilo music generator
Generate music from a text description and save one MP3 with Sonilo. For a first run, keep the supplied ethereal, gentle music prompt, the 60-second duration and the default V0 MP3 quality, then run the Pipe. No media upload is required. To create another track, describe its musical style, instruments, mood and atmosphere in Music prompt, and set Duration from 1 to 360 seconds. Duration is a target rather than a guarantee of exact track length. Under advanced Audio quality, choose V0 variable bitrate or a fixed bitrate of 128k or 320k. This changes the saved MP3 encoding; the output format remains MP3. Generation sends your prompt to the Sonilo service and uses a provider-priced API call. Sonilo currently ignores the hidden seed, so repeated runs are not guaranteed to reproduce the same music.
Seedance text to video
Generate one video from a description of the scene, subject action and camera movement through the Seedance API. No source image is required. For a first run, use the supplied skier prompt or replace it with your own scene. Keep the default Seedance 1.0 Pro model, 720p resolution, 16:9 framing and five-second duration, then run the tool to receive the video. Choose aspect ratio, resolution and duration to shape the output. Duration is specified in whole seconds from 3 to 12. Fixed camera adds a stationary-camera instruction to the prompt but does not guarantee the effect; avoid requesting camera movement when it is enabled. For generated audio, select seedance-1-5-pro-251215 and enable Generate audio. Other listed models ignore that setting. The default model generates without audio. AI watermark requests the provider's “AI generated” watermark. Seed -1 selects a new random seed on each run; a fixed seed controls generation randomization. The advanced Video output section offers file type, codec and quality (CRF) controls. Leave their defaults for automatic output encoding. These settings affect the delivered video encoding, while the generation controls determine its content. This workflow requires an authenticated ComfyUI partner API account with sufficient credits. Generation sends the prompt and generation settings to the provider and uses a paid API call.
LTX image to video
Animate an uploaded starting image into a short, silent video using LTX-Video 2B v0.9.5. For a first run, upload an image and replace the fox example in the positive prompt with a description of your subject, intended movement and camera behavior. Keep the default 768 × 512 dimensions, 97 frames, image strength 0.15, 30 sampling steps, CFG 3 and Euler sampler. At 24 FPS playback, the returned clip lasts about 4.04 seconds. Leave batch size at 1 and output container, codec and CRF at their source defaults. Width and height use multiples of 32 with a minimum of 64 pixels. Frame counts are 9, 17, 25 and so on in increments of 8. Image strength controls starting-image conditioning. The negative prompt describes unwanted visual or motion characteristics. Larger dimensions and frame counts increase processing and memory requirements. Playback FPS changes how quickly the generated frames play and sets duration without adding frames. Conditioning FPS supplies a separate frame-rate cue during generation. Their source defaults are 24 and 25 FPS respectively; set them equal when you want the two settings aligned. Sampling controls include text guidance, initial noise, seed, steps, algorithm and the LTX schedule's base and maximum shifts. Keep their defaults for a first run. Seed -1 requests fresh sampling noise each run; use a fixed seed to compare changes with the same initial noise. Disabling initial noise removes that seed's effect on initial noise selection. Video playback settings offer 8-bit or 10-bit output. The source describes 10-bit as retaining smoother gradients, but player and encoder support varies. Video output settings select automatic encoding or an explicit MP4, MKV or WebM container and H.264 or AV1 codec. Use compatible container and codec choices; CRF interpretation depends on the codec. This graph returns one assembled video. Increasing latent batch size does not create separate video outputs. Only the image loader's IMAGE output is connected; its alpha mask is unused, so uploaded transparency is not preserved. There is no audio input or separate end-frame control.
LTX-2.3 image to video with audio
Animate a source image into one video with generated audio. Upload an image and replace the example prompt with directions for the scene, action, camera movement, dialogue and sounds. The tool generates video and audio together, performs fixed 2× spatial upscaling on the video latent, then refines both streams before returning the video with its soundtrack. For a first run, keep the requested 1280 × 720 size, nominal 5-second duration and 25 FPS. Leave Text-to-video mode and prompt enhancement disabled, and keep the sampling and encoding defaults. The image is center-cropped to the requested aspect ratio; Image fit can instead stretch it. It is then resized to a fixed 1,536-pixel longer edge and compressed before image conditioning. Uploaded alpha is not used as a transparency channel. Width and height also determine the initial half-resolution generation size. Use positive dimensions and duration, and a whole-number frame rate from 1 to 120. Requested frame count is duration × FPS + 1; keep this at or below the audio stage's 1,000-frame limit. Model alignment can affect actual dimensions and duration, so these settings are requests rather than exact output trimming. Image strength controls initial source-image conditioning. Refinement image strength separately controls conditioning reapplied after upscaling. Text-to-video mode bypasses both conditioning stages, so neither strength affects that mode. An uploaded image is still required because its processing path remains connected. If prompt enhancement is enabled, the image can still influence the expanded prompt in Text-to-video mode. Prompt enhancement expands your instructions before video prompt encoding. Leave it disabled to use your written prompt directly. Enhancement tuning and its text LoRA strength apply to expansion; temperature, token filters, penalties and the enhancement seed belong to sampled text generation. Thinking mode only applies if supported by the loaded text model. Initial generation and refinement have independent seeds, CFG guidance, samplers and noise schedules. Use -1 for randomized seeds; fix both generation seeds when comparing settings, and also fix the enhancement seed if using sampled expansion. Negative-prompt influence depends on each stage's CFG and can be limited at the default of 1. Keep the distilled LoRA strength and manual schedules at their supplied values for the first run. Advanced output settings choose bit depth, container, codec and CRF. Start with 8-bit assembly, automatic container and codec selection, and CRF -1. Explicit encoding choices must be compatible with each other and your player; 10-bit output can have compatibility limitations. Audio is generated from the prompt rather than supplied through an audio upload. The model, text encoder, audio decoder and spatial upscaler are fixed parts of this tool.
Wan 2.6 image to video
Animate a source image into a short Wan 2.6 video, with optional automatically generated audio. For a first run, upload the required source image and replace the sample car prompt with motion, camera movement, and visual style suited to your image. Prompts support English and Chinese. Adjust the car-specific negative prompt to describe unwanted features for your subject. Start with the default 720P resolution and 5-second duration, then run the Pipe to receive the video in Sweet Tea. Resolution choices are 480P, 720P, and 1080P. Duration choices are 5, 10, and 15 seconds. References to 4K and 60fps in the sample prompt are creative text, not output settings; this Pipe has no FPS control. Prompt enhancement is enabled by default and allows AI assistance to expand your prompt. Shot type selects single for one continuous shot or multi for multiple shots with cuts; this choice only takes effect while prompt enhancement is enabled. Generated audio is enabled by default. Turn it off when you do not want automatic audio. This Pipe does not accept an uploaded audio track. Seed -1 selects a fresh random seed each run; a nonnegative seed controls generation randomization. Watermark enables the provider's AI-generated watermark. Advanced video output settings let you choose file type and codec; both default to auto. Generation runs through the Wan partner API using the fixed wan2.6-i2v model. The source image uses LoadImage's RGB image output; uploaded transparency is not retained.
Negative Embedding for Deliberate
A negative prompt embedding for Deliberate V2. Use in your negative prompt to make the image look better. Do not use as a regular prompt.
