Répertoire des ressources
Trouvez ce dont vous avez besoin pour votre flux de travail.
Guides: techniques, setup, and troubleshootingRecherchez Sweet Tea, Pipes, des flux de travail externes, des modèles, des packs de nœuds et des guides dans un catalogue unique lié aux preuves.
Commencez plutôt avec un flux de travail ↗Résultats de la recherche
Toutes les ressources
19 affiché of 42 605
Wan 2.2 speech-driven video
Generate a speech-driven video from a reference image, a speech recording and a scene prompt. Separated vocals condition generation; the trimmed original recording supplies the soundtrack. Additional outputs preview the prepared reference, separated vocals and instruments. First run: upload both required files, replace the example positive prompt to match your reference and desired motion, and retain the source defaults: 960 × 528 fill/crop framing, 49 frames per chunk, 5 steps, CFG 1, uni_pc sampling, simple scheduling and 16 FPS. The negative prompt describes unwanted content and also feeds negative-attention guidance. Seeds of -1 randomize each run; use fixed seeds when you want to retain the noise choices. Width, height and framing prepare the reference and determine generation dimensions. Use positive dimensions divisible by 16. Stretch, crop and pad offer different ways to fit the reference; keep-proportion mode can change the resulting dimensions. Frames per chunk follows the 1 + 4n convention, such as 49. It does not choose the number of chunks: the graph generates an initial segment and one continuation, while five further source examples remain bypassed and are not selectable here. Advanced settings retain independent initial and continuation denoising, generation batch size, three stacked LoRA strengths, sampling shift and negative-attention guidance strength, mixing and clipping. Start with their authored defaults. Batch size 1 is the single-video recipe, not a chunk-count setting. The Lightning HIGH and LOW LoRAs are applied in series, followed by Lightx2v; they are not alternative model modes. FPS changes playback timing without generating more frames or retiming audio. The starting frame offset defaults to 2 and changes the final frame selection without trimming the soundtrack. Both controls can affect synchronization. Audio trimming remains fixed at source positions 0–6; the supplied evidence does not establish those positions' units or boundary behavior. Output duration is not automatically matched to the recording. The video output defaults to MP4 with NVENC H.264 and requires a compatible encoding runtime. Model and encoder selections retain the authored recipe. Uploaded transparency is not preserved through the RGB reference-image path.
Flux Klein crop inpaint
Replace a selected region with a Flux 2 Klein generated crop and blend it onto the original RGB image. First run: upload the image, provide a manual mask through the image input's mask workflow, and replace the example Edit instruction with your desired change. Leave Use mask enabled, Automatic mask disabled and Use native crop size enabled. Set Object to select to an object present in your image: SAM2's image output feeds the crop even in manual-mask mode. Run and inspect the edited image, processed mask and input crop preview. The processed mask preview is before cutoff and final feathering, not the final blending mask. For text-detected selection, keep Use mask enabled, enable Automatic mask and name the object to select. Detection threshold, blur, offset and inversion adjust this selection. Disable Use mask for whole-image image-to-image processing. Neither mask switch bypasses SAM2 processing. The crop supplies both the starting latent and a latent reference. Denoise controls change strength throughout that crop; the selected mask controls cropping and final compositing, not a diffusion noise mask. Crop padding adds surrounding context in source-image pixels. Mask coverage adjusts the resized mask, Mask cutoff thresholds it, and Composite edge applies final expansion and feathering. Inverting a processed or composite mask does not change the earlier crop selection. Native crop size uses the crop dimensions as resize targets, with processing dimensions subject to the resize stage's divisibility setting. Disabling it stretches the crop to a square whose width and height equal Sampling resolution in pixels; this is not a maximum-side limit. The graph resizes the generated image back to the crop dimensions but does not explicitly route a correspondingly restored processed mask into compositing. Inspect alignment in this mode. Large native crops can increase memory use. Advanced controls include seed, sampling algorithm, schedule, steps, Flux guidance and resolution-dependent sampling shifts. Seed -1 requests fresh randomization; a fixed seed reuses the noise choice. Desaturation reduces generated-crop saturation before blending, with a selectable grayscale conversion method. The source suggests 0.05–0.20 for saturation mismatch, but this does not guarantee brightness or color matching. Uploaded transparency is not preserved as output alpha.
Hamunaptra Style
This is a Style embedding which turns everything in your image towards an Ancient Egyptian style! Think Pyramids, the river Nile, and Pharaohs! Simply throw Style-Hamunaptra on to the end of your prompt (or the...
FLUX.2 Klein Base 9B text to image
Create an image or a batch of images from text using the fixed FLUX.2 Klein Base 9B model. Describe the subject, composition, lighting and style in the positive prompt; optionally list unwanted elements in the negative prompt. No uploaded image is required. For a first run, keep the 1024 × 1024 size, one image, 20 steps, CFG 5 and Euler sampler. The supplied motorcycle prompt is a ready-to-run example. Width and height jointly set the latent image size and the resolution-aware sampling schedule; use multiples of 16. Larger dimensions and batches increase memory requirements. Advanced controls let you tune guidance, steps, sampler and batch size. Higher CFG increases the influence of prompt conditioning and can change the visual balance; more steps increase sampling work without guaranteeing a better result. Seed -1 randomizes each run; use a fixed seed to reuse the initial noise while comparing other settings. The runtime requires flux-2-klein-base-9b-fp8.safetensors, qwen_3_8b_fp8mixed.safetensors and full_encoder_small_decoder.safetensors. The source graph uses the Base model with its supplied sampling settings; it does not provide an image-editing mode or a selectable distilled-model preset.
HiDream O1 image studio
Create an image batch with HiDream O1 Image Dev by editing an uploaded reference or generating from a text description. For a first run, upload a source image, keep Image edit enabled, replace the example Positive prompt with your requested change, leave Refine prompt off, and run with the existing defaults. Results are returned through Sweet Tea's image output; this graph does not save files through ComfyUI SaveImage. An upload is required by this interface in both modes. With Image edit enabled, the original upload supplies reference conditioning. Output width and height follow the upload, rounded down to multiples of 32. Use dimensions of at least 64 pixels and keep the calculated canvas within 4096 pixels per dimension. To calculate a different edit canvas, enable the currently bypassed Resize edit output stage and set Edit megapixels, measured in millions of pixels. Aspect ratio follows the upload. This stage changes canvas sizing only; reference conditioning and prompt refinement still receive the original upload. Resize method affects the intermediate sizing copy, rather than directly filtering the generated result. Disable Image edit for text generation and describe the complete desired image in Positive prompt. Text width and Text height specify pixels, in multiples of 32 from 64 to 4096. Text batch count and Edit batch count independently select the number of results for their respective modes. Text mode does not use the upload as image-generator reference conditioning, but the interface still requires it. Refine prompt optionally uses Gemma 4 and the upload to expand your request into an English visual description, including in text mode. It may add interpretation or inferred details; leave it off when exact wording matters. Refinement tuning controls text sampling, its separate seed and the response token limit. Temperature and token filters apply when Sample refinement text is on. An insufficient token limit or malformed JSON response can prevent prompt extraction. Refinement thinking requests model-supported thinking behavior; false is the source configuration for JSON extraction. The starting image settings are 28 steps, normal scheduler, CFG 1 and Add noise enabled. CFG changes conditioning influence; at 1, conventional classifier-free negative guidance is inactive. Denoise changes the sampling schedule from an empty latent in both modes, rather than directly controlling preservation of uploaded pixels. Advanced Sampling noise tuning exposes the initial and final per-step noise multipliers and clipping threshold. Image Seed and Refinement seed control separate random processes: -1 randomizes each run, while fixed seeds help compare settings. This tool accepts one reference image. Uploaded alpha is not carried through the RGB reference path. Edit canvas dimensions are rounded rather than clamped, so check the resulting size when using a very small image or an extreme aspect ratio.
Grok Imagine image generator
Generate one or more images from a text prompt using Grok Imagine. Describe the subject, composition, lighting, style, and any desired text. No uploaded image is required. For a first run, use the supplied example or replace it with your own prompt. Keep grok-imagine-image-quality, the square 1:1 aspect ratio, 1K resolution, and one image, then run the Pipe. Returned images appear in Sweet Tea results; the output stage does not permanently save files. The model field selects the Grok Imagine model. Aspect ratio sets the width-to-height shape, while resolution selects a size tier supported by that model. Keep the supplied values unless you know another combination is supported; the supplied evidence does not enumerate alternative model, aspect-ratio, or resolution values. Number of images requests 1–10 results from the same prompt. Seed controls whether the provider node re-runs, rather than making images reproducible. Leave it at -1 to randomize on each run. A fixed seed does not guarantee the same image. Generation sends the prompt to the external Grok provider and uses a paid provider call.
Character camera views
Upload a character image and describe the camera view you want. This Pipe uses Qwen Image Edit 2511 with the multiple-angles LoRA to generate one image per run. The uploaded image supplies both reference conditioning and the initial image latent. For a first run, upload your image, keep the close-up instruction and leave Fast mode enabled. Replace the instruction to request another view. The eight source camera instructions are shown below; leading whitespace and trailing line breaks inside the code blocks reproduce the source strings. Close-up: ``` Turn the camera to a close-up. ``` 45 degrees right: ``` Rotate the camera 45 degrees to the right. ``` Aerial: ``` Turn the camera to an aerial view. ``` Wide-angle: ``` Turn the camera to a wide-angle lens. ``` Low-angle: ``` Turn the camera to a low-angle view. ``` 90 degrees right: ``` Rotate the camera 90 degrees to the right. ``` 90 degrees left: ``` Rotate the camera 90 degrees to the left. ``` 45 degrees left: ``` Rotate the camera 45 degrees to the left. ``` Fast mode coordinates the Lightning LoRA, Fast steps and Fast CFG through one source boolean. Its defaults are 4 steps and CFG 1. Disable Fast mode to use the model without Lightning and the independent Standard steps and Standard CFG defaults of 40 and 4. Each mode's tuning fields affect only that mode. Keep Euler and the simple scheduler for the source setup. Negative prompt guidance has little or no effect at CFG 1. Denoise defaults to 1 to allow substantial viewpoint changes. Reducing it retains more of the initial image structure but can prevent the requested camera change. Camera LoRA strength adjusts the multiple-angles adaptation in both modes; Lightning strength applies only in Fast mode. These are generated interpretations of a single image, not measured 3D rotations. Hidden surfaces and character details may be invented or changed, and requested degree values are instructions rather than geometric guarantees. The source automatically scales the input before generation and offers no independent output dimensions. Uploaded transparency is not preserved by this RGB-only path. Run again with another instruction to create additional views.
Color balance
Create a color-graded image by adjusting red, green and blue separately in shadows, midtones and highlights. For a first run, upload a source image and run with the saved settings to preview the source's stylized grade. These defaults include strong shadow and highlight shifts rather than a neutral starting point. To start without a color shift, set all nine color adjustments to 0, then adjust the tonal region you want to change and run again. Each adjustment uses a unitless -100 to 100 scale. Positive red, green or blue values add that channel; negative values shift toward cyan, magenta or yellow respectively. Tonal regions are weighted by the original image's HSL lightness and overlap, so they are not hard selections. The nine adjustments combine in one grade. In advanced controls, Preserve luminosity is enabled by default. It restores original HSL lightness after the shifted RGB values are clamped. Disable it to allow the color shifts to alter lightness too. This preserves HSL lightness, not every measure of perceived brightness; strong adjustments can clip colors. The result has the source image's dimensions. Uploaded transparency is not preserved because the image loader's separate alpha mask is not used. This tool applies only color balance; it does not combine the other adjustment tools in this family.
Style Sylva Magic
This is a Textual Inversion trained on one image with DreamArtist, my second publicly released inversion. :)
LTX-2 video with generated audio
Generate a video with accompanying generated audio from a text description. Describe subjects, appearance, actions over time, camera movement, and desired sounds, dialogue, or singing. No uploaded image or audio is required. The selected output combines the decoded imagery and audio into one video; MP4 is the default container. For a first run, keep the supplied puppet scene, 121 frames, both frame rates at 24, and the sampling defaults. This gives approximately five seconds. Replace Scene and sound with your own description when ready. Negative prompt lists unwanted characteristics and supplies conditioning to both sampling stages. Width and height are nominal output targets. The initial pass runs at half those dimensions, followed by a fixed 2× spatial latent upscale and a short distilled refinement pass. Internal quantization can change the final size, including the supplied 1280×720 target. For aligned dimensions, use multiples of 64, at least 128 pixels in each dimension. Use frame counts of 8n+1, such as 121, within the audio stage's supported range of 1–1000. Video FPS controls conditioning and playback; Audio FPS independently determines audio duration from the shared frame count. When changing frame rate, set both to the same integer from 1 to 120. Changing one does not update the other. Optional dolly-left camera guidance is disabled by default in both stages. Enable Initial camera guidance and Refinement camera guidance together to apply it across both passes, starting with their supplied strengths of 1. Their toggles and strengths are independent. A prompt requesting a stationary camera may compete with enabled dolly-left guidance. Advanced controls separate initial generation from refinement. Initial steps and schedule shifts affect the first pass. Refinement uses its own sampler, CFG, distilled adapter strength, and explicit noise-level sequence; Initial steps does not alter that sequence. Keep the supplied refinement settings for the first run. Each pass has its own seed: -1 randomizes on each run; set both seeds to fixed values when comparing settings. Generated speech, singing, and timing require review and may not match the prompt exactly. Video output exposes container, codec, and compression controls. Leave MP4, codec auto, and CRF -1 for the first run. Video precision selects 8-bit or 10-bit output; 10-bit requires compatible downstream encoding and playback. The fixed setup uses ltx-2-19b-dev-fp8.safetensors in checkpoints, gemma_3_12B_it_fp4_mixed.safetensors in text_encoders, ltx-2-19b-distilled-lora-384.safetensors in loras, and ltx-2-spatial-upscaler-x2-1.0.safetensors in latent_upscale_models. Optional camera guidance uses ltx-2-19b-lora-camera-control-dolly-left.safetensors in loras. All five assets are marked MISSING_MODEL in the supplied dependency evidence; local availability has not been established. The checkpoint, distilled adapter, and upscaler have supplied sources under https://huggingface.co/Lightricks/LTX-2/resolve/main/ using their filenames. The text encoder source is https://huggingface.co/Comfy-Org/ltx-2/resolve/main/split_files/text_encoders/gemma_3_12B_it_fp4_mixed.safetensors. The optional camera adapter source is https://huggingface.co/Lightricks/LTX-2-19b-LoRA-Camera-Control-Dolly-Left/resolve/main/ltx-2-19b-lora-camera-control-dolly-left.safetensors.
Mage-Flow image edit
Create edited images from one uploaded reference and written instructions using the fixed Mage-Flow-Edit model. Upload your source image, then describe what to keep and what to change. An instruction such as “Keep the subject and pose; replace the background with a grassy prairie” is a starting point. Preservation requests guide generation but do not guarantee unchanged pixels. For a first run, leave Resize reference enabled at 1 megapixel, Match reference size enabled, Number of results at 1, and sampling at 30 steps with instruction guidance 5. Leave Negative prompt blank unless you want to describe unwanted content or visual traits. Run the Pipe to receive the edited image results. Reference resizing retains the reference aspect ratio and aligns dimensions to multiples of 16. When Match reference size is enabled, output dimensions follow this processed reference, so Reference megapixels also affects output size. Turning Resize reference off passes the original image dimensions through. Turning Match reference size off uses Width and Height instead; reference preprocessing remains independent. Use positive manual dimensions divisible by 16. With resizing off and size matching on, use a source image whose dimensions are divisible by 16. Advanced controls include reference pixel area and interpolation, seed, sampling steps, instruction guidance, sampler, scheduler, and denoising amount. Seed -1 selects fresh randomization for each run; a fixed seed helps compare settings using the same sampling seed. The source sampling defaults are Euler, Simple, and denoising amount 1. Denoising amount acts on the edit encoder's latent output and is not established as a source-image preservation strength. This Pipe accepts one reference image and returns the requested number of edited images. It has no mask-based edit boundary, and uploaded transparency is not preserved: only the loader's IMAGE output is connected.
Style Swirl Magic
This is my fifth Textual Inversion, called Style-SwirlMagic. This model was trained on a single image using DreamArtist. As the name suggests, it adds magical swirls to images, making it a great tool for creating RPG...
Swap a product in a character photo
Create an edited photo that replaces the product a person or character is holding with a product from a second photo. Upload the subject photo as image 1 and a clear replacement product photo as image 2. The product reference is resized and padded to the subject photo’s dimensions, then both images are sent to Gemini in that order. This uses a paid Gemini provider call and sends both images and the instructions to that provider. The generated image is returned for preview; this output transport does not permanently save it. For a first run, keep the supplied swap instructions, automatic aspect ratio, 2K resolution and default Gemini model. Add instructions about the subject, pose, lighting, background or product details you want retained. Model selection offers gemini-3-pro-image-preview and Nano Banana 2 (Gemini 3.1 Flash Image). Automatic aspect ratio follows the subject image; an explicit ratio requests a different output shape. Resolution offers 1K, 2K and 4K; 2K and 4K use the provider’s native upscaler. Reference preparation provides advanced white or black padding and resampling controls for the product reference. System instructions control Gemini’s foundational behavior; keep the image-generation directive and put the specific edit in Swap instructions. Seed -1 randomizes each run. A fixed seed requests best-effort consistency and does not guarantee identical results. IMAGE requests image-only generation; IMAGE+TEXT also requests accompanying text, but only the image is returned by this Pipe. This is a generative edit rather than a masked paste. Inspect the result for changes to hands, packaging, lettering and surrounding details. Uploaded alpha transparency is not preserved through the connected RGB image paths.
CyberRealistic Negative - SD 1.5
Join The Tinkerer on Whop. Membership gets you early releases, private tools and a bunch of extra stuff.
Reimagine an image with Grok and refinement
Generate an upscaled, refined reinterpretation of a reference image with Grok and Z-Image Turbo. Upload a reference image and set the style direction and detail instructions. QwenVL describes the reference; your style direction, that caption and your detail instructions form one positive prompt used by both generation and refinement. The uploaded image reaches Grok through its caption only, so identity, layout and details may change. For a first run, replace the fashion and mossy-rock prompt text with instructions for your subject. Keep one image, square aspect ratio, 1K generation resolution, 1 base megapixel, post-upscale scale 0.5 and refinement strength 0.4. The result is the refined image batch. Prompt phrases such as “8k resolution” do not set actual output dimensions. Choose the Grok generation model from the available image, quality and pro variants. Aspect ratio sets the generated composition's proportions, and image count selects 1–10 images that each continue through upscaling and refinement. Generation resolution selects 1K or 2K for the initial Grok image; the later sizing controls determine refinement dimensions. The local Z-Image Turbo refinement model and RealESRGAN 4× upscaler remain fixed. Sizing first fits each generated image to the base megapixel budget, applies the fixed 4× upscaler, then multiplies width and height by the post-upscale scale before refinement. At 1 base megapixel and scale 0.5, final area is approximately 4 megapixels, with twice the base image's width and height, subject to pixel rounding. This is not necessarily twice the upload or Grok output. Scale 1 retains the full 4× base dimensions. Increasing either sizing control increases refinement work. Base and final resize methods choose interpolation at their respective resizing stages. Lower refinement strength retains more of the generated image; higher values permit larger changes guided by the shared positive prompt. Refinement steps, CFG, sampler and scheduler adjust local sampling. The negative prompt applies only to refinement, and its influence depends on CFG; the supplied CFG default is 1. Advanced sampling shift changes the refinement noise schedule and defaults to 3. Generation and refinement always run; there is no standalone upscale or refinement mode. Uploaded transparency is not preserved. Grok requires access to its provider service and remains nondeterministic regardless of its re-run seed. Use -1 for a fresh random seed value each run. The separate refinement seed controls local sampling noise; fixing it does not make the upstream Grok image reproducible.
Wan 2.1 trajectory-guided video
Produces a silent video from a starting image, a motion prompt and point trajectories using the fixed Wan 2.1 ATI model. For a first run, upload your starting image and write a positive prompt describing its subject, intended action and camera movement. Review the negative prompt, whose source default is a Chinese-language exclusion list. Annotate motion paths for that same image with the source-recommended Trajectory Annotation Tool at https://comfyui-wiki.github.io/Trajectory-Annotation-Tool/ and paste its exported JSON into Trajectory JSON. The prompt and trajectory fields start empty because the source examples describe a particular image. Both must be supplied for your composition. Trajectory JSON is an outer array of tracks, each containing an ordered sequence of objects with numeric x and y coordinates. Use the annotation tool's export; the supplied evidence does not establish how arbitrary track lengths are resampled. Check trajectory alignment whenever you change the image composition or output dimensions. Initially retain 720 × 480 pixels, 81 frames, 16 FPS, 20 steps and CFG 3, giving approximately 5.1 seconds of playback. Width and height use multiples of 16; frame counts follow 4n + 1. More frames change the generated sequence length. FPS changes playback timing without generating additional frames. Image guidance optionally adds visual conditioning from the same uploaded image. Disabling it leaves the required starting-image and trajectory conditioning in place. Guidance crop affects only this optional visual conditioning and has no effect when image guidance is disabled. Sampling controls include seed, steps, CFG, sampler, scheduler and denoise. Seed -1 requests fresh random noise; a fixed seed reuses sampling noise. Denoise is not a direct source-image preservation slider because the video latent comes from the conditioning stage. Start with the supplied sampling defaults, including denoise 1 and sampling shift near 8. Keep batch size 1 for the documented single-video use. Trajectory temperature and top-k remain available, but their detailed effects are not documented in the supplied source evidence; initially retain 220 and 10. Video playback controls set FPS and bit depth. Start with 8-bit; 10-bit requires compatible encoding and playback. Video encoding controls select the output container, codec and CRF quality setting; initially retain auto, auto and -1. These settings affect delivery rather than the generated motion. The model and supporting encoders are fixed, and this workflow has no audio input.
Anima depth-guided illustration
Upload a reference image with the spatial layout you want to guide, then describe the new scene and style in the positive prompt. The reference is resized, Depth Anything 3 estimates its depth, and Anima generates a fresh image using that depth as guidance. Outputs include the generated image batch and the grayscale depth preview. For a first run, keep Output size at 1 megapixel, Depth strength at 1, and Turbo mode off. Output dimensions are derived from the depth preview rather than independent width and height fields. The source image supplies spatial guidance, not an encoded starting image: this tool does not preserve its exact colors, texture, identity, or transparency. Turbo mode selects the fixed Anima Turbo LoRA and switches to its separate sampling settings: 8 steps and CFG 1 by default, versus 30 steps and CFG 4 in standard mode. Each mode's steps and CFG remain independently tunable. Turbo LoRA strength only matters when Turbo mode is enabled. Adjust Depth strength to change the influence of the reference layout. Depth guidance start and end specify fractions of the sampling process and should remain ordered from start to end. Depth normalization and sky clipping change the map used both for the preview and generation. Keep the renderer on grayscale depth and the estimator in mono mode for the fixed model and depth patch. Increasing depth-estimation resolution can recover finer structure at greater memory and processing cost; it does not independently enlarge the final image.
0001SoftRealistic Negative and Positive Embeddings
Negative embedding, which I created for myself - I was too lazy to manually write a negative prompt every time. Works very well with my model 0001SoftRealistic - v154xxx | Stable Diffusion Checkpoint | Civitai , but...
Async's Negative
This is Negative Embedding for SD1.5, Not for SDXL . The concept was to improve quality, such as EasyNegative and veryBadImageNegative .
