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
19 angezeigt of 42.632
HappyHorse text to video
Generate a 3–15 second video from an English or Chinese text prompt with HappyHorse. Configured ComfyUI partner API access is required; no uploaded media is needed. For a first run, use the supplied floral fashion prompt or describe your own scene, action, visual style, and camera movement. Keep HappyHorse 1.1, 720P, 16:9, and a 5-second duration, then run the tool to receive the video in Sweet Tea. Choose 720P or 1080P and a duration from 3 to 15 whole seconds. Requests such as “8K” or “60 fps” in the prompt do not override the resolution control; there is no frame-rate control, separate negative prompt, or reference-image input. HappyHorse 1.0 is an alternative model. When selecting it, choose 16:9, 9:16, 1:1, 4:3, or 3:4. HappyHorse 1.1 also supports 21:9, 9:21, 5:4, and 4:5. Check the aspect ratio whenever you switch models. Seed -1 requests a fresh random seed each run; a nonnegative seed controls generation variation. Enable Watermark to request the provider's AI-generated watermark; use this control rather than relying on watermark instructions in the prompt. Advanced Video output settings offer automatic formatting or MP4, MKV, and WebM file types, with automatic, H.264, or AV1 codec choices. Leave both settings on auto for the first run.
Color and edit images with Gemini
Color an illustration or apply prompted image edits with Gemini, or generate an image from text alone. For a first run, leave Source image enabled, upload an illustration, and use “Color this image.” Start with Auto aspect ratio and 1K resolution. Add palette, style and details to preserve or change to your prompt. An upload is required while Source image is enabled. For text-only generation, disable Source image and replace the coloring prompt with a complete description of the image you want. No upload is needed in this mode. Choose an explicit aspect ratio to specify the composition. With a reference image, Auto follows its aspect ratio. Resolution selects a 1K, 2K or 4K target; 2K and 4K use Gemini's native upscaler. Model selects Gemini 3 Pro Image or Nano Banana 2 (Gemini 3.1 Flash Image). Seed -1 randomizes each run; a fixed seed requests best-effort consistency without guaranteeing identical results. System instructions control Gemini's underlying interpretation of the prompt; the default directs it to produce an image. Keep Response format at IMAGE for image-only requests. IMAGE+TEXT also requests text, but this Pipe returns only generated images and does not retain the text response. This workflow sends the prompt and any enabled reference image to the Gemini service and uses a provider-priced request. Access to that service is required. Results are returned through an image preview output. The reference connection carries RGB image data only and does not preserve uploaded transparency.
Image glow
Create an image with a sampled, optionally tinted glow from an uploaded source image. For a first run, upload your image and run with the saved settings: Intensity 2, Radius 29, Glow tint #0a9299, screen blending and Threshold 100. These settings apply a teal-tinted effect rather than a neutral preview. Change intensity or radius and run again to compare. Set Intensity to 0 to preview the source without the effect. Intensity uses a unitless 0 to 100 scale and scales the glow before blending. Radius uses a 0 to 100 scale rather than direct pixel units: the sampling radius is 0.012 × Radius squared in pixels, so the saved value 29 gives about 10 pixels. Sampling includes slight spatial jitter. Radius 0 also disables the effect. Glow tint multiplies the sampled glow's colors. Enter an RGB hex color; #ffffff leaves those colors unchanged. The saved #0a9299 favors teal. Pure black #000000 skips the tint multiplication rather than suppressing the glow. Advanced controls choose brightness selection and blending. Threshold is an inverted 0 to 100 scale: low values favor bright samples, while high values include darker samples. The brightness selection has a soft transition. Choose the exact lowercase blend name add, screen, soft, overlay or lighten; enter it exactly if shown as a text field. Add sums the source and glow, screen combines their inverse values, and lighten keeps the larger value in each channel. Soft and overlay use tonal blends that can darken areas when the sampled glow is weak, so lowering intensity is not necessarily a fade back to the source in those modes. Intensity 0 disables every mode. Strong settings can clip colors to black or white. The active effect also adds a small deterministic dither. 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 glow; the other image adjustments in this family are separate tools.
Kling first and last frame video
Generate one video from a starting image, a motion prompt, and an optional ending image using Kling. Provider access is required; generation sends inputs to Kling and uses a paid provider call. The result is returned as a video preview. For a first run, upload the first and end frames, replace the example prompt to match your images, and keep the five-second, 1080p defaults. The end-frame stage starts enabled and needs an upload. To animate only the starting image, bypass the End frame stage; no end-frame upload is then needed. Describe subject motion, camera movement, desired appearance, and the transition in the motion prompt. The same field accepts desired and unwanted content; there is no separate negative prompt. Duration ranges from 3 to 15 seconds. Resolution offers 720p, 1080p, and 4k. Kling v3 Omni is the default model; Kling Video O1 is also selectable. Generate audio is supported only with Kling v3 Omni, so leave it off with Kling Video O1. Seed -1 randomizes each run; a fixed seed does not make provider results reproducible. The advanced Video output stage provides file-type and codec choices. Keep both on auto for the first run. This interface supports first-frame and first-to-last-frame generation. Storyboards remain disabled because editable segment prompts and durations are absent. Reference-image guidance is not connected in this Pipe.
SwitchX Video Alpha Matte
Generate a separate automatic subject alpha matte video from a trimmed source clip with Beeble SwitchX. SwitchX performs the scene edit as part of this operation; this Pipe returns its matte video, while the edited color video is not included. For a first run, upload a source video, replace the example prompt with your desired scene edit, and keep 720p resolution, start time 0 and duration 5 seconds. Check that the segment contains no more than 240 frames at the source video's frame rate. Leave output file type, codec and CRF at their defaults. The reference image stage is disabled initially. To use it, enable the stage and upload an image to guide the background, lighting or costume. Supply a nonempty prompt of at most 2000 characters, an enabled reference image, or both. Disable the reference stage to use the prompt alone. Start time and duration are measured in seconds. Duration 0 selects the remaining clip, still subject to SwitchX's limits of 240 frames and approximately 2.77 megapixels per frame. Strict duration raises an error when the requested duration is unavailable instead of accepting a shorter segment. Output resolution selects the maximum processing resolution, rather than an exact width and height. Alpha mode remains fixed to Auto for automatic subject isolation. Fill produces no separate matte, and this graph provides no keyframe or custom matte inputs for Select or Custom operation. The returned matte is a separate video, not an edited color video with embedded transparency. Seed -1 randomizes the rerun seed on each run. Identical seeds do not guarantee identical SwitchX results. This workflow sends inputs to the Beeble provider and invokes its paid processing service.
HappyHorse reference-to-video
Create a short video from a character reference image and a scene reference image, guided by a motion prompt. Set resolution, aspect ratio, duration, seed, watermark, and output FPS; choose a HappyHorse 1.1 or 1.0 model. Video file type and codec are also selectable.
Quality Embeddings for Anime models
I made 2 basic quality embeddings as an alternative to the quality tags that are normally used with my checkpoints. They work of course also with any other IL/NAI-based anime checkpoint.
ElevenLabs voice isolation
Reduce background noise in a speech recording with ElevenLabs voice isolation and export the processed audio as an MP3. For a first run, upload your recording under Source audio, keep MP3 quality at V0, run the Pipe, and download the result. Advanced MP3 export settings also offer 128k and 320k; these change MP3 encoding rather than the isolation process. The entire uploaded recording is processed. This Pipe has no trim controls or adjustable isolation strength.
Video green screen
Replace a video's background with a solid green or blue chroma-key backdrop using Bria. Upload the video you want to process, leave Background shade on broadcast_green, and run. Download the resulting video and use your editor's chroma-key tools to remove the colored backdrop for compositing. Background shade offers broadcast_green (#00B140), chroma_green (#00FF00), and blue_screen (#0000FF). Choose a color that contrasts with the foreground. The output has a solid colored background; it is not a transparent video. Seed controls whether Bria processing reruns. Leave -1 to randomize on each run. A fixed seed does not make the provider's results reproducible. Under advanced Video output settings, leave File type and Codec on auto for a first run. Explicit container choices are MP4, MKV, and WebM; codec choices are H.264 and AV1. The supplied evidence does not establish support for every explicit pairing. Audio preservation is also not established by the supplied graph. Processing sends the video to Bria and uses a provider-priced API operation.
Lotus Depth
Upload a source image to generate a Lotus Depth map. Advanced controls let you choose the sampler and schedule, set the step count and denoise amount, and adjust initial sigma. Lotus conditioning is fixed; the default uses one inference step.
Veo video generation
Create one short video with Google Veo from a scene description, optionally guided by an uploaded image and with generated audio. For a first run, replace the example video prompt with your subject, action, setting, camera movement and visual style. Keep the default eight-second, landscape, 720p settings and leave Reference image bypassed for text-only generation. Audio generation is off by default. For image-guided generation, enable Reference image and upload an image. The image guides generation through Veo's optional image input; the prompt and output settings remain available in both modes. Choose 16:9 landscape or 9:16 portrait output and a duration of 4, 6 or 8 seconds. Use the negative prompt for unwanted content or visual traits. Enable Generate audio when wanted and describe intended sounds or speech in the video prompt. People generation lets you allow or block generation of people. Choose a Veo model and resolution. The available resolutions are 720p, 1080p and 4K; 4K is unavailable with Veo 3.1 Lite and Veo 3.0 models. Seed -1 selects a new random seed on each run. A fixed seed is a generation input and does not guarantee identical provider results. Video output contains optional file type and codec choices. Keep both on auto for a first run; these affect the returned video's encoding rather than its generated content. Generation uses an external, provider-priced Veo API. Your prompt and any enabled reference image are sent to the provider.
Love Potion
Do you have a problem with shy subjects in your images? Try this love potion to really break the ice.
Image relight
Create a relit version of an uploaded image using a lighting instruction, Qwen Image Edit, and a dedicated relighting LoRA. For a first run, upload an image and keep the default warm twilight instruction, denoise 1, relighting LoRA strength 1, and the sampling preset: 4 steps, CFG 1, Euler sampler, and simple scheduler. Run the Pipe to receive one edited image. To request another lighting setup, describe direction, time of day, color temperature, softness, and shadows. The relighting trigger is added automatically. The source image is resized before supplying both image conditioning and the initial latent. Output dimensions follow this preprocessing; there are no independent width or height controls. Lower denoise values retain more of the initial image's structure and can limit the lighting change. The default 1 allows a full denoising pass. Objects and details may change along with the light. Uploaded transparency is not preserved because only the loader's RGB image output is connected. Relighting LoRA strength controls the dedicated adapter independently of denoise; 0 removes its contribution while retaining the underlying image-edit model. The negative prompt describes unwanted attributes, but at CFG 1 it has no ordinary classifier-free exclusion effect. Increasing CFG enables its influence and changes sampling behavior. Leave it blank for the source preset. Use seed -1 for a new randomized run, or a fixed seed to compare instructions and settings with the same initial noise. Advanced controls also expose guidance normalization strength and mode, sampling shift, and Lightning adapter strength. Normalization before CFG can amplify the result; the source preset leaves it off. Start with the supplied values and evaluate changes together with CFG, steps, sampler, and scheduler. Removing or changing the Lightning adapter's influence changes the basis of the four-step preset. The workflow requires qwen_image_edit_2509_fp8_e4m3fn.safetensors, qwen_2.5_vl_7b_fp8_scaled.safetensors, qwen_image_vae.safetensors, Qwen-Image-Edit-2509-Relight.safetensors, and Qwen-Image-Edit-2509-Lightning-4steps-V1.0-bf16.safetensors. Installation sources are supplied, but the dependency records mark these artifacts MISSING_MODEL; this review does not establish that they are installed.
Autumn Style
This is a DreamArtist Textual Inversion Style Embedding trained on a single image of fall/autumn like colors. You can get amazing autumn/fall color effects in your images by calling Style-Autumn in the positive...
Krea 2 Turbo image generator
Create an image or batch of images from a text description using Krea 2 Turbo. No uploaded image is required. For a first run, replace the example image prompt, choose an aspect ratio, and keep the defaults: 1 target megapixel, one image, prompt refinement enabled, and Darkbrush disabled. Run the Pipe to receive the generated image batch in Sweet Tea. Prompt refinement uses the local text model with fixed prompt-expansion instructions. Turn it off to use your wording directly. Refinement can change the interpretation of a detailed prompt. Its token limit, sampling settings, seed, optional thinking mode, and model template are available under Refinement tuning. Thinking applies only if the model supports it. Sampling-specific controls apply when refinement sampling is on. Image sampling and prompt refinement have separate seeds. Use -1 for a new random seed each run. To repeat a setup, fix the image seed and either disable refinement or also fix its seed and settings. Enable Darkbrush to apply the fixed style LoRA and append the Style phrase to the direct or refined prompt. LoRA strength changes the model adjustment; the phrase separately changes text conditioning. Setting strength to zero still leaves the phrase appended. Disabling Darkbrush selects the base model and omits the phrase. This interface uses the fixed Darkbrush LoRA and does not select other style LoRAs. Aspect ratio and target megapixels jointly determine pixel dimensions. At the square default, 1 target megapixel is approximately 1024 × 1024. Increasing resolution or image count increases resource use. The selected output returns the whole image batch through Sweet Tea's preview transport; it does not permanently save images through ComfyUI SaveImage. Image sampling starts from an empty latent with 8 steps, CFG 1, Euler sampling, the simple scheduler, and denoise 1 by default. Advanced Image sampling controls change this process. Lower denoise values do not preserve an uploaded image because this workflow has no image input. Negative conditioning is zeroed, so there is no independent negative prompt.
Illustration to realism
Reinterpret an uploaded illustration as a realistic photographic image using Qwen Image Edit and the Anything2Real adapter. Upload an illustration and keep the default editing instruction for a first run, or describe the photographic lighting, materials, setting, and features you want retained. Begin with Denoise at 1, Realism strength at 1, and the supplied 8 steps, CFG 1, Euler sampler, and beta scheduler. The source image supplies both visual conditioning and the initial latent. This is a generative reinterpretation, so details and identity can change. Lower Denoise to retain more of the original structure, potentially leaving more of its illustrated appearance. Realism strength scales the Anything2Real adapter; setting it to 0 or bypassing the Realism adapter stage keeps prompt-driven image editing active without that adapter's influence. Use Seed -1 for a new random variation each run, or a fixed seed to compare changes with the same starting noise. Negative prompt can be left blank. At CFG 1, ordinary negative-prompt guidance has no effect; increasing CFG makes positive-versus-negative guidance meaningful. Steps, CFG, Sampler, and Scheduler remain independent tuning choices; start with the supplied settings rather than assuming larger values improve the result. Advanced settings expose Lightning strength, guidance normalization, and sampling shift. Lightning strength defaults to 1 for the supplied eight-step adapter; 0 removes its influence without changing Steps automatically. Guidance normalization has its own strength and a before-guidance switch: off uses post-guidance attenuation, while on applies unclamped normalization before guidance combination and can amplify the prediction. Sampling shift defaults to 3 and changes the model's noise schedule independently of the Scheduler choice. Output shape derives from the uploaded image through latent encoding; there are no independent resize controls. Uploaded transparency is not preserved because only the loader's RGB image output is connected.
Edit character expressions
Edit a character image with one of 12 preset expressions or a custom expression. Upload an image, choose preset 1–12 or select 13 and enter a custom instruction. Batch mode generates all 12 preset expressions. Set output size and expression strength; advanced controls include editable prompt text and sampling options.
Qwen-Image: Text to image
Generate images from text with Qwen-Image, using standard sampling or an optional Lightning LoRA mode. No source image is required. For a first run, replace the example positive prompt with your subject, composition, lighting and style. Include desired written text in the prompt. Leave the negative prompt empty unless you want to discourage particular elements. Keep 1328 × 1328 pixels, image count 1 and Lightning off. Standard mode defaults to 20 steps and CFG 4. Lightning mode selects Qwen-Image-Lightning-8steps-V1.0.safetensors and its separate defaults of 8 steps and CFG 1. The toggle switches the model, steps and guidance together. Each mode has independently editable steps and CFG; only the active mode's settings affect the result. Lightning strength applies only in Lightning mode. Negative-prompt influence depends on CFG; Lightning's default CFG is 1. Width and height are pixels in multiples of 16. Set them together to choose the aspect ratio; 1664 × 928 is landscape and 928 × 1664 is portrait. Increasing dimensions or image count increases memory use. Seed -1 selects a fresh random seed each run; use a fixed seed when comparing settings. Sampler, scheduler, denoise and sampling shift are shared by both modes. The supplied defaults are Euler, simple, full denoising at 1 and sampling shift approximately 3.1. This graph starts from an empty latent and does not edit uploaded images. The fixed model files must be available: qwen_image_fp8_e4m3fn.safetensors in diffusion_models, qwen_2.5_vl_7b_fp8_scaled.safetensors in text_encoders and qwen_image_vae.safetensors in vae. Their source is https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main/split_files. Lightning additionally uses Qwen-Image-Lightning-8steps-V1.0.safetensors in loras, supplied at https://huggingface.co/lightx2v/Qwen-Image-Lightning/resolve/main/Qwen-Image-Lightning-8steps-V1.0.safetensors. Dependency evidence currently marks these files as missing; ensure they are available before running the corresponding setup.
HiDream I1 Full text to image
Generate images from text using the fixed HiDream I1 Full FP8 model. No uploaded image is required. Describe the subject, composition, lighting and style in the positive prompt; use the negative prompt for unwanted attributes. For a first run, replace the example positive prompt or run it as supplied. Keep 1024 × 1024 pixels, image count 1 and the supplied advanced settings: 50 steps, CFG 5, uni_pc sampler, simple scheduler, sampling shift 3 and denoise 1. Run the Pipe to receive the generated image batch. Width and height must be multiples of 16; set them together to choose the aspect ratio. Image count controls how many images are generated per run. Increasing dimensions or image count increases resource use. Seed -1 randomizes each run; use a fixed seed to reuse the initial noise when comparing settings. Advanced sampling exposes steps, CFG, sampler, scheduler, denoise and sampling shift. CFG controls prompt guidance; at CFG 1 the negative prompt has no effect. Sampler and scheduler change the sampling process, while sampling shift adjusts the model's noise schedule. Denoise defaults to 1; lowering it reduces the effective noise schedule. Because generation starts from an empty latent, it does not preserve an existing image. The model and text encoders are fixed. Dev and Fast models are not selectable modes, and this Pipe does not offer image-to-image input.
