ComfyUI-KJNodes

This node functions by cloning the input model and setting an optimized attention override function within its model options.

Speeds up VAE decode/encode with fused Triton norm+SiLU kernels and channels_last conv layout. Supported VAEs (auto-detected): Wan 2.1/2.2 video VAEs incl. Qwen-Image (RMSNorm, ~1.4x/1.15x), KL image VAEs such as Flux/Flux2, SDXL and SD1.5 (GroupNorm, ~1.6-1.8x at 2048px), and LTXV/LTX2 video VAEs (PixelNorm; timestep-conditioned decoder blocks get norm-only fusion). Other architectures are not supported. Applied as object patches on a cloned patcher, so it only exists while this VAE is loaded.

Per-token prompt weighting for Krea2 (K2) via attention value scaling. Use (word:-1) to remove a concept, (word:1.5) to emphasize one -- works through the Qwen3-VL encoder where normal weighting doesn't. Outputs the patched model + conditioning; set the sampler CFG to 1.0.

Adds a dedicated live-preview frame on this node, with overridable max resolution. Default ComfyUI preview caps at 512px; this node sends its own preview straight to a DOM widget on the node so pixel-space models (Chroma Radiance, ZImage, HiDream-O1, …) can be previewed at full sampler resolution.

This node technically supports multiple upsampling algorithms (such as bilinear, Lanczos, etc.) for image scaling and can simultaneously process images and their associated mask channels.

It analyzes the color distribution of the target image and the reference image, applying the color style of the reference image to the target image using multiple algorithms.

This node is implemented based on the `color-matcher` library and provides multiple color transfer algorithms, including MKL, HM, Reinhard, MVGD, and their hybrid methods.

Visual prompt builder for Ideogram 4's structured JSON caption format. Drag on the canvas to draw regions; select a region to set its type (obj/text), description, text, and color palette. Set the background and optional style fields as widgets. Outputs the assembled caption JSON string. bbox is normalized to a 0-1000 grid as [ymin, xmin, ymax, xmax]; width/height set the canvas aspect ratio. Canvas controls: - Drag: draw a new region - Ctrl/Cmd-drag: force-draw a new region even on top of an existing one - Click: select a region · Alt-click: cycle overlapping regions - Double-click: edit the description inline - Right-click: region list (select / delete / duplicate / reorder, top = front) - Del / Backspace: remove the selected region - Ctrl/Cmd + C / V / D: copy / paste / duplicate the selected region - bbox fields (px / out) next to obj/text are editable Color swatches: - Click: edit · Drag: reorder · Right-click: remove - Hover + Ctrl/Cmd + C / V: copy / paste the hex - "+": add a color (uses the clipboard color if it is one) Toolbar: - Live: use the live sampling preview as the background (and grab the final result) - Grab BG / Clear BG: use the last generated image as the background - brightness slider, token estimate, and Copy / Paste / Clear all

EXPERIMENTAL! Activates custom sageattention to reduce peak VRAM usage, overrides the attention mode. Requires latest sageattention version.

EXPERIMENTAL! Activates custom sageattention on the WanVideo self-attention to reduce peak VRAM usage, overrides the attention mode. Requires latest sageattention version.

EXPERIMENTAL AND MAY CHANGE THE MODEL OUTPUT!! Chunks feedforward activations to reduce peak VRAM usage.

Returns the frames captured by Model Preview Override during the most recent sampling. Wire 'model' from Model Preview Override (the same one feeding the sampler) and 'after_sample' from after the sampler (LATENT/IMAGE) to enforce correct execution order.

https://github.com/ChenDarYen/Normalized-Attention-Guidance

This node not only supports loading standard VAE model files but also specifically integrates support for various TAESD (Tiny AutoEncoder for Stable Diffusion) approximation models, including dedicated versions for images (e.g., taesd, taesdxl) and videos (e.g., taehv, lighttaew2_2).

This node technically first converts the input image data into a tensor format suitable for model processing, then batches the images according to the user-defined `per_batch` value and feeds them into the specified super-resolution model for processing.

This node provides fine-grained control over weight data types and computational data types when loading models, supporting various precision formats from fp8 to fp32.

Its core functionality is to load a main GGUF model, and optionally load and fuse a second additional model, such as a VACE module.

This node technically works by cloning the input model and binding `ON_PRE_RUN` and `ON_CLEANUP` callback functions to it.

Experimental node for patching attention to use flash attention, without the silent SDPA fallback the ComfyUI default does. Patches the attention of the model passing through this node; to disable, bypass or disconnect this node. Requires the flash_attn library to be installed.

This node technically processes by converting the input image tensor into a PIL image object and then encoding it in memory.

This node supports reading image files in JPG, JPEG, PNG, WEBP, and TGA formats.

Overrides the LTX2 preview sampling preview function, temporary measure until previews are in comfy core

This node technically extracts width and height values by parsing the user-selected preset string, and calls ComfyUI's built-in `EmptyLatentImage` function to generate a latent noise tensor of the specified dimensions and batch size.

This node technically works by dynamically creating input ports