KreativeSuite

Self-contained video refinement, upscaling, and image-layer decomposition nodes for ComfyUI.

KreativeSuite turns substantial media workflows into focused nodes with nested model selection, compact controls, prompt assistance, progress reporting, and ComfyUI-native previews.

Important

KreativeSuite is an experimental preview. The nodes are already usable, but model interfaces in newly released ComfyUI builds can change. Save a copy of important workflows before updating ComfyUI or the suite.

Hover a parameter in any public node for concise guidance. Nested Settings fields also include a ? help marker with the same explanation.

Included nodes

Kreative LTX 2.5 Video Upscaler

A self-contained LTX 2.5 video-to-video refinement and spatial upscaling pipeline.

• Loads a video directly from the node with familiar frame-rate, frame-cap, skip, every-nth-frame, and resize controls.
• Preserves source identity and motion through an adjustable image-input strength.
• Offers output presets for 1080p, 1440p, 4K, 2x source, or a custom resolution.
• Includes the official two-stage sigma schedule as the official scheduler.
• Also exposes simple, normal, beta, and sgm_uniform for more generative finishing passes.
• Bundles optional adaptive contrast and sharpening with an adjustable amount.
• Trims frame counts to the LTX-compatible 8n+1 grid and keeps audio aligned to the selected source interval.
• Reports sampling progress and an estimated completion time.
• Plays the selected source video in an in-node preview before generation.
• Uses QwenVL to inspect uniformly sampled frames across the complete timeline, preserving shot changes, ordered actions, and camera motion instead of captioning only one frame.
• Uses a separate Gemma prompt-writer model to convert that temporal report into a faithful LTX refinement prompt.
• Lets you edit the generated prompt before refining, or run analysis and refinement as one combined action.
• Supports Analyse Prompt, Refine Video, and Analyse + Refine modes.
• Uses ComfyUI's native latent-preview system.

Kreative H3 Video Refiner

A self-contained MiniMax H3 refinement pipeline for an existing video.

• Loads the H3 transformer, optional Turbo LoRA, text encoder, video VAE, and audio VAE from a nested settings panel.
• Exposes Turbo LoRA strength on the node; the tested starting point is strength 0.8 with 2 Euler/simple steps.
• Can analyse representative frames with a local vision model through Ollama.
• Lets you review and edit the generated prompt before committing to the refinement pass.
• Supports Analyse Prompt, Refine Video, and Analyse + Refine modes.
• Loads the source video and its audio directly from the node.
• Plays the selected source video in an in-node preview before refinement.
• Keeps output sizing, aspect ratio, target megapixels, denoise, and sampling controls on the main interface.
• Keeps source timing, clip-length policy, model selection, and video/audio time-shift controls inside Settings.
• Keeps only the optional reference-images socket as an external input.

Kreative Stable Layers

A self-contained Qwen Image Layered pipeline that separates one image into four ordered RGBA layers.

• Accepts a single ComfyUI IMAGE input and returns layers_rgba, alpha_masks, and a reconstructed composite.
• Runs the trained decomposition stage at a fixed 640-pixel maximum dimension; there is no editable generation-resolution control.
• Reads the original source dimensions automatically and returns every layer at that exact size.
• Uses premultiplied-alpha bicubic enlargement to keep RGB and transparency edges aligned during the high-resolution second stage.
• Returns the four layers as an ordered batch, back to front, ready for layer-aware saving or further processing.
• Accepts both the original Stability Labs PEFT adapter and a ComfyUI-converted adapter.
• Keeps the Qwen Image Layered model, Stable Layers adapter, Qwen 2.5 VL text encoder, and layered RGBA VAE inside Settings.

Installation

From the custom_nodes directory in your ComfyUI installation:

git clone https://github.com/formulake/KreativeSuite.git

Restart ComfyUI, then refresh the browser interface. The public nodes appear under the KreativeSuite category.

ComfyUI Manager installs the packages listed in requirements.txt automatically. For a manual installation, run this with the Python interpreter used by your ComfyUI installation:

python -m pip install -r requirements.txt

To update later:

cd KreativeSuite
git pull

KreativeSuite does not download model weights. Install each required file through ComfyUI Manager or place it manually in the matching ComfyUI model folder.

Required models

The filenames below are the defaults used by the nodes. Compatible alternatives can be selected under Gear / Settings.

LTX 2.5

Component	Default file	ComfyUI folder
Distilled transformer	ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors	models/diffusion_models
Video VAE	ltx-2.5-video-vae-bf16.safetensors	models/vae
Audio VAE	ltx-2.5-audio-vae-bf16.safetensors	models/vae
LTX conditioning encoder	gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors	models/text_encoders
Gemma prompt writer	gemma4_e4b_it_fp8_scaled.safetensors	models/text_encoders
x2 latent upscaler	ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors	models/latent_upscale_models

MiniMax H3

Component	Default file	ComfyUI folder
H3 transformer	minimax_h3_ref2va_pruned_int8_convrot.safetensors	models/diffusion_models
Turbo LoRA (optional)	minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors	models/loras
H3 text encoder	qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors	models/text_encoders
Video VAE	minimax_h3_video_vae_fp16.safetensors	models/vae
Audio VAE	minimax_h3_audio_vae_fp32.safetensors	models/vae

Stable Layers

Component	Example file	ComfyUI folder
Qwen Image Layered model	qwen_image_layered_bf16.safetensors or a compatible INT8 ConvRot build	models/diffusion_models
Stable Layers adapter	stable_layers_comfy.safetensors or the original PEFT adapter	models/loras
Qwen 2.5 VL text encoder	qwen_2.5_vl_7b_fp8_scaled.safetensors	models/text_encoders
Layered RGBA VAE	qwen_image_layered_vae.safetensors	models/vae

The base model is available from Comfy-Org/Qwen-Image-Layered_ComfyUI, and the adapter is published at StabilityLabs/Stable-Layers. The compatible community INT8 ConvRot build tested during development is available at appmana/qwen-image-layered-int8convrot.

LTX prompt analysis requires ComfyUI-QwenVL-Mod and a compatible QwenVL model. The tested default is the local Qwen2.5-VL-3B-Instruct; QwenVL uniformly samples 32 frames across the complete loaded timeline by default, then the selected standalone Gemma model writes the LTX prompt. The analyser model, frame count, device, and Gemma writer are editable in Settings.

H3 prompt analysis also needs Ollama and a vision-capable model. The default is qwen3-vl:32b; both the model and endpoint are editable in the node settings. H3 refinement itself does not require Ollama when you provide the prompt manually.

The optional acceleration adapter is available from LightX2V/Minimax-H3-Turbo. When that exact file is installed, new H3 nodes select it automatically and start at strength 0.8, 2 steps, and Euler/simple sampling. Select None or set its strength to 0 to use the unpatched base model. If the adapter is not installed, new nodes retain the base-model defaults of 16 steps and res_multistep.

Model weights are not included in this repository and remain subject to their respective licences and access terms.

Using the LTX upscaler

1. Add Kreative LTX 2.5 Video Upscaler.
2. Choose an existing input video or press Load video....
3. Set the output resolution and source-preservation strength. 0.3 reproduces the tuned source workflow.
4. Leave the scheduler on official for the supplied two-stage custom sigma sequence and zero-denoise finishing stage.
5. Enable adaptive contrast only when the source benefits from stronger local separation.
6. Open Gear / Settings to select the LTX models, QwenVL video analyser, Gemma prompt writer, analysis frame count, analysis device, and source-video handling.
7. Choose Analyse Prompt to inspect uniformly sampled frames from the complete source timeline without starting the diffusion stages.
8. Review or edit the returned prompt, then choose Refine Video. Use Analyse + Refine when no manual review is needed.
9. Connect video to a compatible save-video node, or use refined_frames and audio independently. The prompt_used output exposes the exact prompt used for the refinement.

Analysis-only loads the source, QwenVL video analyser, and standalone Gemma prompt writer, but skips the LTX transformer, VAEs, latent upscaler, and sampling stages. Refine Video always uses the text currently visible in the prompt field without regenerating it.

Selecting a scheduler other than official turns the finishing stage into a generative pass governed by steps and denoise. That can add detail, but it can also alter faces, identity, texture, or motion.

Using the H3 refiner

1. Add Kreative H3 Video Refiner.
2. Choose an existing input video or press Load video.... The node loads both its frames and source audio.
3. Optionally connect reference images.
4. Set output sizing, aspect ratio, target megapixels, denoise, and sampling on the main interface.
5. Open Gear / Settings and select the H3 model components and optional Turbo LoRA.
6. With the Turbo LoRA installed, start with strength 0.8, 2 steps, and Euler/simple sampling. Increase the steps or disable the LoRA when comparing against the base model.
7. Choose Analyse Prompt to generate a description without starting inference.
8. Edit the returned prompt and creative direction if needed.
9. Choose Refine Video and queue the node again.

All H3 model components and the optional Turbo LoRA are selected in Gear / Settings and loaded internally when the workflow runs. The LoRA is applied to the transformer only; selecting None or strength 0 bypasses it cleanly.

Using Stable Layers

1. Add Kreative Stable Layers and connect one source image.
2. Open Settings and select the four required model components.
3. Describe the desired decomposition in the prompt, or keep the concise default for a general foreground-to-background separation.
4. Queue the node. Decomposition runs internally at the model's trained 640-pixel size.
5. Use layers_rgba for the four transparent layers, alpha_masks for their masks, or composite to inspect the reconstructed image.

The 640-pixel inference size is intentionally fixed for predictable speed. Layer enlargement is a separate alpha-aware stage that takes its target width and height from the connected source image, so all outputs remain registered to the original canvas.

Source preview, latent preview and ETA

The selected input video plays directly inside either node. Playback starts muted so browsers can autoplay it; use the preview controls to seek, pause, or enable audio.

Generation previews use ComfyUI's global latent-preview setting. Select a preview method in ComfyUI settings; turning global latent previews off also turns them off for KreativeSuite. The ETA becomes meaningful after the first sampling step and applies to the currently active diffusion stage.

Compatibility

• A current ComfyUI build with native LTX 2.5, MiniMax H3, and Qwen Image Layered support is required.
• The suite relies on ComfyUI's installed PyTorch, NumPy, Pillow, and PyAV packages; it does not create a separate Python environment.
• Windows is the primary tested platform. The node code itself does not intentionally depend on Windows-only paths.

Troubleshooting

• A model is missing from a selector: place it in the folder shown above, then restart ComfyUI or refresh the model list.
• The source preview does not play: confirm that the selected video uses a codec supported by your browser. MP4 with H.264 video is the most broadly compatible choice.
• No latent preview appears during generation: enable a latent-preview method in ComfyUI settings.
• Ollama analysis cannot connect: confirm that Ollama is running, the selected vision model is installed, and the endpoint in Settings is reachable.
• LTX temporal analysis is unavailable: install and enable ComfyUI-QwenVL-Mod, select an installed QwenVL model, and restart ComfyUI. The tested default is local Qwen2.5-VL 3B.
• The output duration is wrong: verify source FPS, forced FPS, frame skipping, and every-nth-frame settings. The LTX node also trims to 8n+1 frames.
• Identity or motion drifts: increase source preservation, reduce finishing-stage denoise, or return to the official scheduler.
• Stable Layers model or adapter does not attach: confirm that the base model is a Qwen Image Layered build and that the selected adapter is either the original Stability Labs PEFT file or a compatible ComfyUI conversion.
• Transparent layer edges look misaligned: confirm that downstream nodes have not resized RGB and alpha independently; use the node's source-sized layers_rgba output directly.

When reporting a problem, include the ComfyUI version, the complete error text, selected model filenames, input dimensions/frame count, and the saved workflow JSON. Do not attach model weights.

Development

The repository intentionally keeps the public UI compact while its expanded internal graph remains cacheable and compatible with ComfyUI's normal model offload and VRAM management.

Basic syntax checks:

python -m py_compile __init__.py alpha_utils.py stable_layers.py h3_refiner.py ltx25_upscaler.py muse_h3_video_refine.py
node --check js/kreative_suite.js
python -m unittest tests.test_alpha_utils

Contributions and reproducible bug reports are welcome through GitHub issues and pull requests.

Acknowledgements

KreativeSuite builds on ComfyUI's native model support and publicly documented work for LTX Video, MiniMax H3, Qwen Image Layered, and Stability Labs Stable Layers. The H3 refinement design was informed by the work of Muse Collective, while the adaptive-contrast implementation is adapted from ComfyUI Essentials. See THIRD_PARTY_NOTICES.md for the included third-party notice.

KreativeSuite is an independent community project and is not an official product of ComfyUI, Lightricks, MiniMax, Ollama, Qwen, Stability AI, or Muse Collective.