Finden Sie, was Sie für Ihren Workflow benötigen.

Clutter

Fill your background with a lot of stuff with a quick prompt. Gives any scene that "lived in" look and feel.

FLUX.2 [dev]1024x10248GB VRAM

Seedream layer separation

Separate an uploaded image into requested visual layers with generated transparency. Upload a source image and describe which elements should belong together, such as a character, background, lettering, or decorative graphics. The connected image outputs are combined with their corresponding generated masks and returned as one image batch. For a first run, replace the example layer instruction with elements present in your image, keep output resolution at 1K, and leave Crop layers off. The instruction requests layer organization; it does not guarantee a fixed number of output images. Enable Crop layers to request cropped layers, or leave it off to request the uncropped layout for compositing. Output resolution offers auto, 1K, 1.5K, and 2K; independent width and height controls are unavailable. Prompt optimization offers the provider's standard and fast modes. Watermark enables the provider's watermark option. Seed -1 requests a fresh randomized seed on each run; a fixed seed is supplied to the provider without guaranteeing identical results. Transparency comes from the generated separation masks. The source image's alpha-mask output is not connected, so uploaded transparency is not preserved as an input mask. Results are returned through Sweet Tea's image output; this graph does not permanently save PNG files.

Grok image edit

Edit a source image with a written instruction using Grok. For a first run, upload one source image, describe the changes you want and the features to retain, and keep the supplied defaults: grok-imagine-image, 1K resolution, one result and request seed -1. The example instruction requests anime styling while retaining identity. The output returns the edited images. Model selects among grok-imagine-image, grok-imagine-image-quality and grok-imagine-image-pro. Resolution selects a 1K or 2K output request. Image count requests 1 to 10 edited results from the same reference and instruction. Request seed controls whether the editing request reruns. Keep -1 for randomization or change a nonnegative seed to trigger another request. Results remain nondeterministic even with a fixed seed. Aspect ratio remains automatic. The source allows manual aspect ratios only with multiple connected reference images, while this tool connects one. Increasing image count changes the number of results, not the number of references. A source image is required; this tool has no text-only mode, mask-based editing input or denoise-strength control. Uploaded alpha transparency is not preserved because only the loader's RGB image output is connected.

Minimax H3 INT8/INT6/INT4 ConvRot

Minimax H3 INT8 / INT6 / INT4 ConvRot

FLUX.2 dev image generation

Generate images from a text prompt using the fixed FLUX.2 dev model, with an optional Turbo LoRA mode. No source image is required, and there is no negative prompt input. For a first run, replace the example positive prompt with your subject, composition, lighting and style. Keep output and scheduler dimensions at 1024 × 1024, image count at 1, guidance at 4, sampler at Euler and Turbo mode disabled. Standard mode uses the separate Standard steps setting, initially 20. Run the Pipe to receive the generated image batch. Width and height set output dimensions in pixels and must be multiples of 16. Whenever you change them, set Scheduler width and Scheduler height to the same values: these independent fields control the noise schedule and are not synchronized automatically. Image count sets batch size; larger batches require more memory. Enable Turbo mode to select the fixed Turbo LoRA and Turbo steps together. Turbo steps defaults to 8. Each mode uses only its own step setting; enter an integer from 1 to 4096. Turbo strength affects only the Turbo model branch. Setting strength to 0 leaves Turbo steps selected; disable Turbo mode to return to Standard mode. Advanced controls include prompt guidance, seed, both step counts, Turbo strength and sampler. Seed -1 chooses fresh random noise each run; use a fixed seed when comparing settings. Sampler selection applies in both modes and can change the result at the same seed and step count. The fixed model stack uses flux2_dev_fp8mixed.safetensors in diffusion_models, mistral_3_small_flux2_bf16.safetensors in text_encoders, full_encoder_small_decoder.safetensors in vae, and Flux_2-Turbo-LoRA_comfyui.safetensors in loras. All four are marked MISSING_MODEL in the supplied dependency evidence; ensure they are available before running. Source-provided downloads are https://huggingface.co/Comfy-Org/flux2-dev/resolve/main/split_files/diffusion_models/flux2_dev_fp8mixed.safetensors, https://huggingface.co/Comfy-Org/flux2-dev/resolve/main/split_files/text_encoders/mistral_3_small_flux2_bf16.safetensors, https://huggingface.co/black-forest-labs/FLUX.2-small-decoder/resolve/main/full_encoder_small_decoder.safetensors, and https://huggingface.co/ByteZSzn/Flux.2-Turbo-ComfyUI/resolve/main/Flux_2-Turbo-LoRA_comfyui.safetensors. Model files are fixed by this Pipe rather than selectable controls.

Inzaniak's Lazy Textual Inversions

You can now get all my LoRAs in one single package if you subscribe to my Patreon (including those not on CivitAI):

Amazing Embeddings - fcNegative + fcPortrait suite

fcNeg is a Negative Embedding. You just install it in the embedding folder and then type fcNeg in the negative prompt. I don't know why it shows two trigger words. Ignore "fcneg-neg"

UpscalingFLUX.1 [dev]1024x10248GB VRAM

FLUX image upscale

Enlarge an uploaded image using the workflow's fixed FLUX/Dev and AnimeSharp upscale preset. For your first run, upload a source image and run the Pipe. The image passes through the Ultimate Upscale stage, and the saved upscaled image is returned as the result. The preset selects the FLUX/Dev fp8 bundle and 4x-AnimeSharp.pth. This interface offers only the source-image upload: there is no editable prompt, seed override, model selector, LoRA selector or optional processing mode. Processing settings remain at their source values. The upscaler's filename does not establish the final enlargement factor, and the supplied evidence does not establish an exact output size. No separate face-detailing stage is exposed. Running requires ComfyUI-UmeAiRT-Toolkit and the preset's model assets. The supplied dependency records leave the toolkit, FLUX/Dev bundle and AnimeSharp asset unresolved; automatic installation is not established.

SwitchX Video Edit

Create one edited video clip using SwitchX, guided by a prompt and optional reference image. For a first run, upload a source video, replace the example prompt with your desired look, and keep auto editing mode and 720p resolution. Start time defaults to 0 seconds and duration to 5 seconds. Shorten the duration if needed to stay within the 240-frame limit at your source video's frame rate. Enter auto to isolate the main subject automatically and edit the surrounding scene while preserving unmasked source pixels. Enter fill to regenerate the whole frame while preserving geometry; the subject's original pixels are not preserved, so this mode can change the subject's look or costume. Select and custom require first-frame keyframe or per-frame matte inputs unavailable in this Pipe; use auto or fill. The reference-image stage is disabled by default. Enable it and upload an image to guide the background, lighting or costume. Disable it for prompt-only editing. Provide a nonempty prompt or an enabled reference image; at least one is required. Prompts are limited to 2,000 characters. Start time and duration are in seconds. Duration 0 selects the remaining video, still subject to SwitchX's limits of 240 frames and approximately 2.77 megapixels per frame. Five seconds may exceed the frame limit for some videos. Enable strict duration to raise an error when the requested duration cannot be extracted; otherwise a shorter available clip is allowed. Enter 720p or 1080p for the maximum output resolution. Seed -1 randomizes on each run. Seed controls provider re-running rather than reproducibility; results remain nondeterministic with a fixed seed. The advanced video-output section provides file type and codec choices; keep both automatic for the first run. This Pipe returns the edited video, not SwitchX's separate alpha video. SwitchX uses an external Beeble provider call with provider billing.

SwitchX1024x10248GB VRAM

UpscalingAnima1024x10248GB VRAM

Anima illustration with upscale and finishing

Generate an Anima illustration from prompts and return an enlarged, finished image batch. For a first run, describe your subject, scene, composition, and style in Positive prompt. The two additional prompt fields append optional details. Leave Negative prompt empty unless you want to specify unwanted content or visual traits. No uploaded image is needed. Choose the initial Size preset, Image count, and Enlargement factor. The default 832 × 1216 portrait preset and 2× enlargement produce 1664 × 2432 images. The factor multiplies both width and height. Leave the sampling defaults unchanged for a first run. Advanced generation controls provide the initial seed, steps, prompt guidance, sampler, scheduler, and denoise. Keep Generation denoise at 1 for the active empty-latent generation path. Generation and tile seeds are independent; -1 requests fresh randomization for that stage on each run. Fix both seeds when comparing changes without deliberately randomizing either stage. Finishing first applies the configured model upscaler, resizes to the chosen enlargement factor, and refines the enlarged image with tiled diffusion. Tile denoise controls how much is redrawn. Tile steps, scheduler, and sampler use the connected custom sampling nodes; the retained fallback fields are not the operative controls. Tile dimensions follow the original generated image dimensions. Tile padding supplies surrounding context in pixels, and Tile mask blur softens blending boundaries in pixels. Noise cleanup and Blur cleanup run after tiled refinement. Their strength choices range from Light to Maximum. Each cleanup stage can be bypassed independently. Strong cleanup can remove texture, so compare a lower strength or a bypassed stage if the result loses detail. Color correction offers Auto, Preset, Manual, and Colorize modes. Choose Preset to use Color preset and Preset variation; choose Manual for direct tonal tuning. Tint applies only in Manual mode. White balance shifts between cooler blue and warmer orange, while Green–magenta balance is a separate adjustment. Skin-tone preservation requires AI analysis. For Colorize mode, choose a colorization method; Force colorization overrides content detection that would otherwise skip art or anime. Sepia blend applies to vintage colorization. Bypass Color correction to return the preceding cleanup result. No color-reference image is connected, and palette and histogram outputs are not returned. The source's hand, face, and eye refinement passes and their detection chains are bypassed, with no Pipe switch to enable them. Their retained settings do not affect runs or provide anatomy repair. The uploaded-image and custom-width/height branches are also disabled; normal runs use the size preset. No LoRA entries are configured. The Pipe returns one final image collection through Sweet Tea's preview output rather than permanently saving through the original Image Saver. Larger batches, enlargement factors, and sampling workloads increase resource use.

CyberRealistic Negative - SDXL

Join The Tinkerer on Whop. Membership gets you early releases, private tools and a bunch of extra stuff.

TextualInversionSDXL 1.0

ACE-Step1024x10248GB VRAM

ACE Step 1.5 music generator

Generate an MP3 music track from a music description and lyrics using ACE Step 1.5 XL Turbo. Describe the genre, instruments, mood, and vocal style, and supply lyrics with optional section markers. Adjust tempo, time signature, lyrics language, and musical key. Output duration and conditioning duration are independent controls in seconds; set them to match when changing track length. One shared seed controls conditioning and audio sampling; -1 selects randomization. Audio code generation is enabled by default and can be toggled, with separate guidance and sampling controls. Additional controls expose batch size, audio sampling settings, sampling shift, and MP3 encoding quality.

Qwen-Image1024x10248GB VRAM

Qwen Image canvas outpainting

Extend an uploaded image on selected sides with prompted scene continuation using Qwen Image and inpainting ControlNet. The output is a resized, composited image with an automatically created padding mask. For a first run, upload an image and replace the example scene continuation prompt with content that fits its subject, lighting and style. The defaults add 200 pixels on the left, top and bottom and none on the right, before resizing the padded canvas to a longest edge of 1536 pixels. Set at least one extension amount above zero. Keep Lightning enabled at strength 1, with 4 steps, CFG 1, denoise 1, Euler and Simple for the initial run. Seed -1 randomizes each run; use a fixed seed to compare settings. Each extension is measured in original-image pixels before resizing. The padded image and its padding mask share the longest-edge target and resize together, retaining aspect ratio. Use a positive target size. The original image is also resampled, so preservation outside the generated region refers to the resized working image rather than the original file's exact pixels. No separate mask upload is needed. The region is derived from directional padding; uploaded alpha and loader edit masks are unused, and output transparency is not preserved. Boundary feathering acts before resizing. Mask expansion then grows or contracts the region in resized-mask pixels: positive values create overlap into the original image, while negative values contract it. Tapered corners affect nonzero expansion. Blur radius controls the Gaussian neighborhood and sigma independently controls its spread. The processed mask guides ControlNet and the final blend; softened boundaries mix original and generated pixels. The padded canvas is encoded and sampled without a separate latent noise mask. Final compositing determines where the generated result is visible, retaining the resized source outside the processed mask. Denoise controls how much of the encoded padded canvas is changed during sampling; lower values retain more initial structure. Control strength adjusts image-and-mask guidance. Guidance start and end are sampling-progress fractions from 0 to 1; keep start below end. The default interval covers the full sampling process. A negative prompt can discourage unwanted content, but has little or no effect at CFG 1. Sampler, scheduler and model-specific sampling shift are independent controls; retain Euler, Simple and shift 3.1 for the first run. Lightning is an optional adapter and starts enabled. Its toggle does not update steps or CFG. When disabling it, manually use the source baseline of 20 steps and CFG 2.5; when enabling it again at strength 1, return to 4 steps and CFG 1. For replacing a selected region inside an existing canvas using an uploaded edit mask, use the separate masked inpainting tool.

Wan 2.1 VACE 14B1024x10248GB VRAM

Recraft image generation from style references

Generate one to six images from a text prompt using a Recraft style created from four uploaded reference images. All four references are required and contribute to the shared style; they guide visual treatment rather than providing an image-editing canvas. For a first run, upload four references with a consistent look, replace the example image prompt with your subject and composition, and keep Output size at 1024x1024 pixels and Image count at 1. Choose realistic_image for a photographic base or digital_illustration for an illustrated base. Output size must be a Recraft-supported width-by-height value. Use Negative prompt to describe unwanted elements, or leave it blank. Image count supports 1–6 images with the same prompt and reference style. Leave Seed at -1 for a randomized re-run seed. A fixed seed controls re-run behavior but does not make Recraft results reproducible; generation remains nondeterministic. This tool requires access to the Recraft partner API. Reference images and generation inputs are sent to the service, and style creation and image generation use provider-priced calls. Style creation remains part of the connected generation path. Reusing a manually supplied style ID is not an available mode in this Pipe.

Wan 2.1 video outpainting

Expand a video's borders with prompt-guided content using Wan 2.1 VACE, producing a silent generated video. For a first run, upload a source video and enter a scene prompt describing its content, the surroundings to add and any desired motion. Keep the initial padding and generation settings, then adjust the padding on each edge to suit your composition. Padding is measured in source-frame pixels and creates the control canvas and outpainting mask; output width and height independently set the generated resolution. Choose output dimensions of at least 16 pixels in multiples of 16. Choose a positive frame count of 4n + 1, such as the default 81. These requirements are documented here because the dimension and frame-count fields have broad numeric ranges. Output FPS sets playback speed independently of source FPS. The default 81 frames at 16 FPS lasts about 5.1 seconds. Source audio is not carried into the result. Frame count is selected separately from the uploaded video's length; this tool does not promise to preserve the original duration or reproduce source pixels exactly. Advanced controls include unwanted-content prompting, VACE control strength, mask feathering and sampling settings. Seed -1 selects randomization; use a fixed nonnegative seed to reuse sampling noise with otherwise matching settings. Keep generation batch size at 1 for a single-video first run. The fixed CausVid LoRA is enabled by default, paired with 3 sampling steps and CFG 1. Its model and text encoder strengths are separately adjustable; text encoder strength only has an effect if the LoRA contains applicable weights. To use the supported bypass mode, bypass the CausVid stage and change sampling to the documented 20 steps and CFG 6. Bypassing does not change those sampling settings automatically. Video encoding offers automatic or explicit container and codec selection, plus CRF quality control; -1 retains the transport's default quality. Video creation defaults to 8-bit. The source describes optional 10-bit creation as retaining smoother gradients, with possible player or downstream compatibility limitations.

Flux Kontext multi-reference editing

Create an edited image from a primary image and up to two additional references using Flux.1 Kontext Pro. Each run also returns a preview of the assembled reference image sent to generation. For a first run, upload the required Primary image and replace the example Edit prompt with your desired change. Identify the referenced subjects and state what should remain unchanged. Both additional references are enabled initially: upload an image for each enabled reference, or disable any unused reference. Disable both for single-image editing; either additional reference can also be used independently. Keep the default layout initially. The second reference is placed above the primary image, and the third is placed to the right of that assembly. Advanced layout controls set each reference's position, joining-edge size matching, gap in pixels, and gap color. Third reference placement is relative to the primary image alone when the second reference is disabled. These controls change the assembled input, not the generated image's aspect ratio. Set Aspect ratio to the desired output width-to-height ratio, between 1:4 and 4:1. Generation controls include guidance strength and step count; their source defaults are 3 and 50. Seed -1 randomizes each run; a fixed seed reuses the seed setting. Prompt expansion lets the provider elaborate your instructions, but prevents exact repeatability even with a fixed seed. The reference preview accompanies generation; it is not a separate preview-only mode. This workflow uses a partner API and requires configured access and sufficient credits. Uploaded transparency is not preserved because only the image outputs, not the alpha masks, enter the reference assembly.

FLUX1024x10248GB VRAM

ACE-Step1024x10248GB VRAM

ACE-Step 1.5 song generator

Generate an MP3 song from a musical description, lyrics and musical settings using the fixed ACE-Step 1.5 XL SFT model. For a first run, keep the supplied English neo-soul description and lyrics, or replace both with your own. Describe genre, instruments, vocals, mood and arrangement in Music description. Enter the words to sing in Lyrics; optional section markers such as [Verse] and [Chorus] communicate arrangement. Keep Generate audio codes enabled, Batch size at 1, both duration fields at 120 seconds and the sampling defaults, then generate the MP3. Output length sets the audio canvas length in seconds. Conditioning duration tells the text-conditioning stage the intended song length. These controls do not synchronize automatically: normally change both together. Tempo is in beats per minute. Musical key, time signature and language provide musical direction rather than guarantees of exact performance. Time signature accepts the source codes 2, 3, 4 or 6, rather than arbitrary fractions. Language uses the supplied supported codes; en means English. E minor is the supplied musical-key value; other accepted key choices are not enumerated in the available evidence. Generate audio codes controls the preliminary language-model stage. Audio-code guidance, temperature, top-p, top-k and min-p apply when that stage is enabled. Disabling it leaves the subsequent audio diffusion stage active; this Pipe does not provide an audio-reference input. Audio-code guidance and Diffusion guidance control separate stages. One shared seed feeds both stages. Use -1 for randomization or a fixed nonnegative integer to reuse a seed. Advanced audio diffusion controls include steps, sampler, scheduler and denoising strength. Full denoising at 1 is the supplied setting for generation from an empty audio latent. Sampling shift adjusts the model noise schedule and defaults to 3. Batch size controls the number of audio latents generated with the shared song settings. Decoded audio is saved as MP3 with V0 variable quality, 128k or 320k encoding. Negative conditioning is automatically zeroed, so there is no editable negative prompt. This is text-and-lyrics generation, with no uploaded audio reference or audio-editing mode.

Wan 2.2 controlled video · Standard

Generate a silent video guided by uploaded control footage and a scene-and-motion prompt using Wan 2.2 14B Fun Control. This standard branch uses high-noise and low-noise models across a coordinated 20-step schedule; the separate four-step sibling uses Lightx2v LoRAs. For a first run, upload a control video and replace the saved scene prompt. Leave Canny edge extraction disabled if the footage already contains prepared control frames. For ordinary footage, enable Canny to extract edges before conditioning generation. Its low and high thresholds apply only while enabled; keep the low threshold below the high threshold. Reference-image conditioning is disabled initially. To use it, upload an image and enable Reference image. You can leave it disabled to generate with the prompt and control video alone. The reference upload supplies the loader's RGB image output, not its alpha mask. Defaults generate 640 × 640 pixels and 81 frames at 16 FPS, approximately 5.1 seconds. Width and height use increments of 16; frame count follows 1 + 4n. Playback FPS changes duration independently of frame count. The result does not inherit source audio or source FPS. Keep batch size at 1 for a first run; this Pipe returns through one video output rather than separate named outputs for batch items. Advanced controls include the negative prompt, random seed, independent prompt guidance for each generation stage, samplers, schedulers and model schedule shifts. Seed -1 randomizes initial noise; a fixed nonnegative seed reuses initial noise with otherwise unchanged settings. Preserve the saved sampling setup initially. If editing schedules, keep both total-step values equal, use coordinated schedulers, and make Initial end step equal Refinement start step within that schedule. The default handoff is step 10 of 20. Refinement continues the initial latent without adding new noise. Video output offers automatic or explicit container and codec choices, plus CRF encoding quality. CRF -1 uses the transport default. Video construction defaults to 8-bit; 10-bit is available but may have player or downstream compatibility limitations. The output is newly generated footage guided by the control frames, rather than a copy of the uploaded clip.

Wan 2.21024x10248GB VRAM

CyberRealistic Positive (Pony)

Join The Tinkerer on Whop. Membership gets you early releases, private tools and a bunch of extra stuff.

Empire Style

This is a DreamArtist Textual Inversion Style Embedding trained on a single image of a Victorian city street, at night. You can get amazingly grand Victorian stone buildings, gas lamps (street lights), elegant...