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 42.618
SDXL Turbo
Generate images from text with the fixed SDXL Turbo checkpoint. Describe your subject, setting, composition, and visual style in Positive prompt; no uploaded image is required. The workflow requires sd_xl_turbo_1.0_fp16.safetensors in ComfyUI's checkpoints directory. For a first run, replace the example positive prompt and keep 512 × 512 pixels, one image, one sampling step, CFG 1, Euler ancestral, Denoise 1, and Add noise enabled. Width and height use pixels in multiples of 8 and determine the aspect ratio. Image count generates a batch of variations. Larger dimensions and batches increase memory use. In advanced Sampling, seed -1 selects new random starting noise each run; use a fixed seed to compare changes from the same starting noise. Advanced controls expose steps, sampler, CFG, denoise, and Add noise. Steps range from 1 to 10; additional steps increase sampling work without guaranteeing a better result. Changing the sampler changes generation behavior. The negative prompt has no effect at CFG 1; raising CFG introduces its influence and changes the source Turbo preset's behavior. Denoise changes the noise schedule used with the empty latent, not the strength of editing an input image. Keep Add noise enabled for normal generation; disabling it removes starting random noise without supplying a source image. This tool produces generated images and does not offer image editing or checkpoint selection.
Qwen-Image depth-guided creation
Generate a prompted Qwen-Image result using the depth structure of an uploaded reference, and receive a preview of the depth map used for guidance. For a first run, upload a reference image whose composition you want to use and replace the supplied example positive prompt with your desired subject, setting, lighting, and style. Leave Lightning acceleration enabled with 4 generation steps, CFG 1, Euler sampling, Simple scheduling, and Denoise 1. Seed -1 requests a new random seed each run; use a fixed seed to compare changes under the same settings. The negative prompt may remain empty and has little effect at CFG 1. The Pipe automatically scales the reference, estimates depth with Lotus, inverts the depth map, and uses that map for InstantX Union ControlNet guidance. Control strength adjusts its influence. Control start and end are fractions of sampling from 0 to 1; keep start below end. Strength 0 removes ControlNet influence but does not change the depth-derived initial latent. The initial generation latent is encoded from the depth map, not the original photograph. Lowering generation Denoise retains more of the depth-map appearance rather than preserving the photograph's colors or details. Output dimensions follow the automatically scaled reference; there are no independent width or height controls. Uploaded alpha is not used by this graph. For base Qwen-Image mode, bypass Lightning acceleration and set generation steps to 20 and CFG to 2.5, following the supplied reference settings. Lightning strength applies only while that stage is enabled. The depth preview remains available in either mode. Advanced depth controls affect Lotus estimation rather than final-image sampling. Depth intensity sets the first sigma; Depth sampler selects the algorithm, and Depth steps, Depth scheduler, and Depth denoise construct the schedule before its first sigma is replaced. Inspect the depth preview when adjusting these settings. Sampling shift affects the Qwen-Image generation schedule independently.
Flux Redux reference blend
Generate new images with Flux.1 dev conditioned by two Flux Redux visual references. For a first run, upload both images, replace the example prompt with your intended scene, and keep reference strengths at 1, output size at 1024 × 1024, image count at 1, and sampling steps at 20. Clear the prompt if you want to omit descriptive text; both reference uploads remain required, even when a reference strength is zero. Each reference has an independent strength and weighting method. The strengths are not percentages and do not need to add up to 1. Multiply scales reference conditioning; attn_bias applies an attention bias. Keep the default multiply method for the first run. The first reference is applied to the text conditioning, then the second reference is added to that combined conditioning. This is not a pixel blend of the uploaded images. Advanced framing controls select center cropping or no center crop independently for each reference before visual encoding. They affect reference preparation, not output dimensions. Set output width and height in pixels, in multiples of 16; image count selects the number generated in one batch. Advanced controls include Flux guidance, seed, steps, sampler, and noise scheduler. Seed -1 selects fresh random noise for each run; use a fixed seed when comparing control changes. The sampling schedule fraction defaults to 1 for the full schedule. Reducing it does not preserve a reference image because generation starts from an empty latent. Sampling calibration exposes maximum shift, base shift, and reference dimensions used to calculate the Flux sampling shift. Those reference dimensions default to 1024 × 1024 and do not automatically follow output size. The result is a newly generated image or image batch, rather than a direct edit of either upload. Exact composition, identity, and pixels are not guaranteed. Uploaded alpha masks are not connected to generation, so transparency is not preserved. There is no negative-prompt branch.
ComfyUI-DareMerge
ComfyUI powertools for SD1.5 and SDXL model merging
Edit or remove masked objects with Flux Fill
Regenerate a masked region of an uploaded image with Flux Fill OneReward, either replacing its content or removing an object. For a first run, upload a PNG whose transparent areas mark the region to edit and whose opaque areas remain outside the edit mask. The core LoadImage node supplies the RGB image and its alpha-derived mask through separate sockets, both connected to inpainting. There is no independent mask upload, and an opaque image supplies no useful edit region. Write instructions describing the desired replacement or the object to remove, then run with the saved sampling settings. Object removal assistance is enabled by default at LoRA strength 1. Keep it enabled for removal. For general masked replacement, bypass Object removal assistance; its strength setting then has no effect. This optional mode uses the same image, mask and sampling path. Advanced controls include Flux prompt guidance, denoise, seed, sampling steps, sampler and noise scheduler. Denoise ranges from 0 to 1 and controls the extent of regeneration. Seed -1 randomizes the noise initialization; a fixed nonnegative seed lets you compare settings with the same initialization. Flux guidance and sampler CFG are independent controls. Negative conditioning is zeroed internally, so this tool has no separate negative prompt. Restrict sampling to mask is enabled by default. Disabling it permits sampling outside the masked region while retaining the mask in inpainting conditioning. Differential Diffusion has its own strength control, separate from denoise and removal assistance. This tool edits the source canvas rather than expanding it and returns a decoded RGB image; uploaded transparency defines the edit mask and is not preserved as output transparency. Masked sampling is not an exact pixel-preservation composite. The graph uses fixed Flux Fill OneReward, text-encoder and VAE assets, plus the removal LoRA when enabled. Available sampler choices do not establish that every combination has been tested.
Wan 2.1 first-to-last frame video
Generate a prompted, silent video transition between a starting image and an ending image using Wan 2.1. For a first run, upload both endpoint images, replace the flower example in Motion prompt with the action or transformation connecting them, and run with the existing defaults: 720 × 1280 pixels, 33 frames and 16 FPS. This produces approximately 2.06 seconds of video. Both uploads are required because each feeds its own visual encoder as well as endpoint conditioning. Images enter through RGB outputs; uploaded transparency is not preserved. The endpoint images guide generative animation rather than exact frame interpolation. Describe subject motion and camera movement explicitly. The supplied negative prompt lists unwanted visual artifacts and static results; edit it in Advanced to suit your scene. Width and height are measured in pixels and must be divisible by 16. Use frame counts of 4n + 1, such as 33, 49 or 81. Duration is approximately frame count divided by FPS. FPS changes playback speed without generating additional frames. Advanced sampling controls include seed, steps, prompt guidance, sampler, scheduler, denoising strength and noise schedule shift. Seed -1 selects a new random seed each run; a fixed seed keeps the initial noise consistent when comparing changes. Start with the supplied sampling defaults. Denoising strength acts on the video latent and is not a direct endpoint-image preservation setting. Start and end guidance crops independently choose none or center for each image's visual encoding. They do not change the separate endpoint-image connections. Sequence batch size defaults to 1; larger batches feed the same video creator and do not yield separate video outputs. Playback also offers 8-bit or 10-bit video; 10-bit compatibility depends on downstream encoding and playback software. Video encoding supports file types auto, mp4, mkv and webm, and codecs auto, h264 and av1. Keep both on auto initially, or choose a compatible container and codec for your delivery needs. This Pipe provides one video output, has no one-image mode or audio input, and uses a fixed Wan model with matching encoders.
Z-Image text to image
Create images from text using the fixed INT8 Z-Image model. No uploaded image is required; decoded images are returned through Sweet Tea's image output. For a first run, use the supplied botanical fashion portrait prompt or replace it with your own subject, setting, composition, lighting, and style. Leave the negative prompt empty unless you want to discourage specific features. Start with one image and the saved sampling settings: 25 steps, prompt guidance (CFG) 4, res_multistep sampler, simple scheduler, denoise fraction 1, and sampling shift 3. Output dimensions are supplied by the upstream resolution selector, whose saved widgets indicate square, approximately one-megapixel settings. This Pipe exposes no resolution controls. Portrait or 8K wording in a prompt does not set image dimensions. Seed -1 randomizes each run. Use a fixed seed when comparing prompt or sampling changes with consistent initial noise. Image count generates a batch with the same prompts and dimensions; larger batches require more memory. More sampling steps increase processing time. Advanced controls include the sampler, scheduler, denoise fraction, and model sampling shift. Denoise starts at 1 for generation from an empty latent; lowering it changes sampling and does not provide image editing. Sampling shift adjusts the model's sampling schedule independently of the sampler and scheduler. The model, text encoder, and VAE are fixed. This workflow has no uploaded-image mode or upscale stage.
Recraft digital illustration
Generate raster digital illustrations from a text prompt with Recraft, an illustration substyle and optional RGB color guidance. For a first run, replace the skateboard example in Positive prompt with your subject, action, composition and desired appearance. Keep the default 2d_art_poster style, 1536×1024-pixel landscape size and one image, then run. No uploaded image is required. Use Negative prompt for unwanted elements, or leave it empty. Illustration style guides the appearance alongside your prompt. Output size specifies width and height in pixels; Image count requests 1–6 images per run. Results appear as raster image previews; this Pipe does not generate SVG output. Color guidance is enabled by default. Its advanced RGB controls start with blue, magenta and yellow-green. Background and palette color supplies both the background guidance and the first palette entry, so changing it affects both. The second and third palette colors are independent. Each RGB channel accepts 0–255. Bypass Color guidance to omit both background and palette guidance; RGB edits have no effect in that mode. Guidance does not guarantee exact color placement or coverage. Generation sends the prompt and generation settings to the Recraft partner API and requires configured access and sufficient credits. Seed -1 randomizes the re-run value for each run. A fixed seed controls re-running or caching but does not make Recraft results reproducible.
Disco Elysium style
Trained on vanilla SD1.5 but works better with art-focused checkpoints like Elldreth's and others from this website. Training dataset contained game portraits, key art, and some cherrypicked fanart.
Dark Fantasy Style
A Dark fantasy style embedding i made using 24 images to train. Turned out pretty good, and you get results just by using the embedding. No trigger words. Also adding things like: scenery, forest, nature, snow,...
Bad Iris Negative (SD 1.5, Embedding)
This negative embedding intends to fix deformed (oval-shaped, weird, etc.) irises in SD 1.5 based models.
Film grain
Create an image with film-style grain from an uploaded source image. For a first run, upload your image and run with the saved settings: grain amount 0.75, size 0.45 and Smooth texture. Adjust amount and size, then run again to compare the texture. Set grain amount to 0 to disable the grain. Grain amount uses a unitless 0 to 1 scale. Grain size controls spatial scale from 0.05 to 3; the source describes 0.3 to 3 as typical. Lower values produce finer texture and higher values produce coarser texture. Choose Smooth for interpolated noise or Grainy for discrete noise, using those exact names if the control appears as a text field. Advanced controls remain useful in both texture modes. Color amount blends monochrome grain at 0 with independent RGB grain at 1. Luminance bias ranges from uniform strength at 0 to strength weighted toward darker pixels at 1, reducing grain in highlights. The noise pattern is deterministic for the same image dimensions and settings; there is no seed control. Added noise is clamped to the available RGB range, so grain can clip at black or white. Output dimensions follow the source image. Uploaded transparency is not preserved because the image loader's separate alpha mask is not used. This tool applies only film grain. The other image adjustments in this family are separate tools and are not combined in this output.
FLUX Canny image generation
The Pipe extracts a Canny edge map from your source image and uses FLUX.1 Canny dev to generate the prompted scene with that structural guidance. Output dimensions follow the source image through conditioning; there are no independent size controls. Edge thresholds tune which outlines guide generation. Guidance controls prompt adherence. Negative conditioning is zeroed internally, so this workflow has no editable negative prompt.
YouTube Thumbnail Generator
Generate a YouTube thumbnail image with Nano Banana 2 from a text prompt, built-in thumbnail art direction and an optional reference image. For a first run, replace the example prompt with your video topic, main subject, emotional reaction and visual hook. Keep 16:9 and 2K, and leave Reference image bypassed for text-only generation. To use visual guidance, enable Reference image and upload an image; an upload is required while that stage is enabled. The reference does not guarantee exact identity or layout preservation. The built-in Thumbnail art direction requests saturated colors, strong contrast, an expressive subject and a simple composition readable at small sizes. If you want lettering, request a short headline; the preset favors 3–5 words. Edit Thumbnail art direction to change this creative brief. Its older Nano Banana Pro wording does not change the actual Nano Banana 2 model. These instructions do not guarantee click-through performance. Aspect ratio selects the frame shape. Auto follows the reference image's ratio when one is supplied; choose 16:9 explicitly for a conventional landscape thumbnail. Resolution offers 1K, 2K or 4K; the provider uses its native upscaler for 2K and 4K. Seed -1 requests a fresh random seed each run. A fixed seed requests best-effort consistency without guaranteeing identical results. Thinking level selects MINIMAL or HIGH provider reasoning. Response mode IMAGE requests an image; IMAGE+TEXT additionally requests response text. This Pipe returns only the generated image, so accompanying text and thought images are not included. Review generated lettering and composition before using the thumbnail.
SCG Emotions Pack - Simple embeds for easily controlling a variety of emotions! (For all 1.5 Models and Merges)
Recently I released my embedding toolkit which has a variety of tools and fixers I've come up with over the past 6 months. I considered adding my emotion embeds to that listing, however I want to keep that specific...
Kling image generation with two references
Generate an image or image series with Kling using a text prompt and two reference images. Upload both reference images and replace the example fashion prompt with a description of your desired result. Explain which subjects, details, or visual qualities to use from each reference. Both references are sent together to the same generation request; this tool provides two required upload slots. Include instructions about unwanted content in the same prompt when needed; there is no separate negative-prompt field. For a first run, keep Kling 3.0 Omni, 1K resolution, and Series amount set to disabled for a single image. Choose the desired aspect ratio. Resolution selects the requested resolution tier, while aspect ratio selects the image proportions. With Kling 3.0 Omni, choose a Series amount from 2 through 9 to request multiple images. Set Series amount to disabled when switching to Kling Image O1, which does not support series generation. Use Seed -1 to randomize on each run. The provider's results remain nondeterministic even with a fixed seed; this control governs rerunning rather than reproducibility. Reference images use the loaders' RGB image outputs. Uploaded transparency is not preserved as an alpha input.
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.
