VELVET VICE — KREA 2 VISION PROMPTER v2.0.5

V2.0.5 – Stable Edit, Registry & Manager Update

VELVET VICE · KREA 2 VISION PROMPTER v2.0.5 is a major workflow and stability update focused on prompt intelligence, reliable image editing, branch isolation, memory handling, ComfyUI Manager integration and a cleaner public release.

The workflow keeps the independent CREATE, Native Edit and Classic Img2Img systems while significantly improving how each branch is executed, detected and managed.

What changed

New KREA Prompt Director v2

• Upgraded to the same Qwen3.5 9B LLM used in the Velvet Vice LTX workflows.

• Improved keyword-to-prompt expansion specifically tuned for Krea 2.

• Better preservation of subjects, actions, camera instructions and important user constraints.

• Adaptive prompt expansion avoids unnecessarily bloated prompts.

• Improved handling of multi-subject scenes, composition, lighting and scene structure.

• Local output validation and cleanup.

• MANUAL mode completely bypasses Ollama.

• ASSISTED mode uses a single LLM call.

• Included Windows installer for the recommended model:

fredrezones55/Qwen3.5-Uncensored-HauhauCS-Aggressive:9b

Prompt-First Memory & Branch Gate

• The Prompt Director finishes before the heavy Krea rendering resources are allowed to resolve.

• Ollama is released after prompt generation whenever possible.

• Heavy CREATE / Native Edit / Classic Img2Img resources remain isolated from inactive branches.

• Inactive branches now use safe ComfyUI ExecutionBlocker behavior instead of leaking invalid None resources.

• Lazy inputs are resolved only when the selected workflow branch actually requires them.

• Prevents inactive VAE, CLIP, MODEL and sampler paths from accidentally executing.

Reworked Native Edit

KREA Native Edit has received a major technical correction.

Two independent Native Edit modes are available:

NATIVE EDIT — ORIGINAL

• Uses the original source format as the edit target.

• Designed for source-faithful editing.

• Best for preserving identity, composition, pose and camera framing.

NATIVE EDIT — CUSTOM

• Uses the source image for visual grounding.

• Generates into a separately defined target resolution / latent.

• Useful when the edited result also needs a different format or aspect ratio.

Additional Native Edit improvements:

• Grounded reference resolution corrected to 768 px.

• Improved Krea2Edit reference wiring.

• Source image, VAE, source latent and target latent are correctly supplied to the edit model path.

• Better distinction between source-preserving editing and free image generation.

Reworked Classic Img2Img

Classic Img2Img remains available as a completely independent image-to-image system.

Unlike Native Edit, Classic Img2Img uses traditional latent denoise strength to determine how strongly the source image is reinterpreted.

Stable public defaults:

• Denoise: 0.40

• Optional creative LoRAs: OFF by default

• SeedVR2: bypassed for the initial base render

This gives users a neutral, source-faithful starting point before enabling stronger reinterpretation or enhancement.

Independent Image Edit Groups

The previous combined Image Edit Studio has been split so Native Edit and Classic Img2Img can be independently bypassed.

New structure:

• 04A — NATIVE EDIT STUDIO • ORIGINAL / CUSTOM

• 04B — CLASSIC IMG2IMG STUDIO • SOURCE / DENOISE

• 04C — IMAGE EDIT BRANCH CONTROL • ENABLE / GUIDE

This prevents Native Edit and Classic Img2Img source paths from unnecessarily running together.

The internal engines also remain independent:

• INTERNAL B — NATIVE EDIT ENGINE

• INTERNAL C — CLASSIC IMG2IMG ENGINE

SeedVR2 Output Isolation

• CREATE, Native Original, Native Custom and Classic Img2Img keep their own output paths.

• New workflow-mode-aware SeedVR2 branch gates prevent inactive finishing paths from being evaluated.

• SeedVR2 receives only the image / model resources belonging to the currently selected workflow mode.

• A dedicated memory handoff can release Krea resources before loading the optional SeedVR2 finishing stage.

New Format & Resolution Selector

Added directly to the CREATE / Text-to-Image branch.

• Choose common aspect ratios instead of manually entering width and height.

• Portrait, landscape, square, widescreen, vertical and ultrawide formats.

• PRESET and MANUAL modes.

• Full manual width / height control remains available.

Render This Stage

Partial execution controls allow selected workflow sections to be tested more easily.

Useful for working independently with:

• Base generation

• Refinement

• Detailers

• SeedVR2 / Final output

This makes troubleshooting and iterative workflow development considerably easier.

Improved PowerLoRA Integration

• Newly created, duplicated or pasted PowerLoRA nodes can inherit the Velvet Vice KREA visual style.

• KREA workflow detection remains isolated from Velvet Vice LTX workflows.

• KREA styling does not overwrite LTX 2.3 / 2.5 graphs.

Memory / Interrupt Improvements

• Improved cleanup after manually interrupted generations.

• Avoids unnecessary aggressive model unloading.

• Designed to reduce situations where large model allocations are unnecessarily pushed into system RAM.

• Prompt resources and render resources are separated whenever possible.

Missing Custom Nodes / Comfy Registry Integration

VELVET VICE KREA is now properly mapped to the Comfy Registry package:

velvet-vice-krea

Velvet Vice workflow nodes include Registry metadata so ComfyUI Manager → Missing Custom Nodes can identify the correct package on clean installations.

This makes sharing the workflow considerably easier because users no longer need to manually identify the Velvet Vice custom-node package.

ComfyUI Manager Install / Uninstall Fix — v2.0.5

v2.0.5 adds dedicated lifecycle support for ComfyUI Manager / Comfy Registry installations.

• Added install.py for legacy-install migration.

• Added uninstall.py for Manager-controlled cleanup.

• Known historical Velvet Vice KREA duplicate folders are automatically handled.

• Read-only file attributes are cleared before Manager removes the active package, reducing Windows deletion / WinError 5 problems.

• The separate VELVET VICE LTX package is explicitly protected and is never modified.

Canonical installation folder:

ComfyUI/custom_nodes/velvet-vice-krea

Legacy KREA folders handled by the cleanup system include:

ComfyUI-Velvet-Vice-KREA
ComfyUI-Velvet-Vice-KREA-main
velvet-vice-krea-main
ComfyUI-ILLUMINATE-AI-KREA

Windows Portable Helpers

The release package includes the original helper BAT files:

• INSTALL_QWEN_3.5_VISION_9B.bat

• run_nvidia_gpu_VELVET_VICE.example.bat

• _INSTALL_AUTOPROMPT_MODEL.bat

The Qwen installer downloads the recommended local model automatically through Ollama.

Emerald + Violet UI

• Updated KREA-specific Velvet Vice color palette.

• Existing Velvet Vice animated node behavior is preserved.

• Theme ownership remains isolated from the separate LTX workflows.

• Layout and spacing were refined throughout the workflow.

Model-Neutral Public Workflow

The public workflow no longer depends on a specific personal Krea checkpoint.

Users are free to select their own compatible Krea 2 model installation.

Model, encoder, VAE and compatible LoRA choices remain under user control.

Independent Workflow Architecture

The four image-generation paths remain logically independent:

ModeSource ImageTargetMain PurposeCREATENoNew latentText-to-ImageNative Edit — OriginalYesOriginal formatPrecise source-preserving editNative Edit — CustomYesCustom target latentGrounded edit into a new formatClassic Img2ImgYesEncoded source latentDenoise-based reinterpretation

Each path has its own rendering and output logic while sharing the central Velvet Vice control system.

Recommended First Test

For the cleanest first test:

1. Select the desired workflow mode.

2. Use MANUAL prompt mode.

3. Keep optional Detailers and creative LoRAs disabled.

4. Keep SeedVR2 bypassed.

5. Confirm the base output first.

6. Enable enhancement stages afterward.

For Classic Img2Img, start with denoise 0.40.

For source-faithful edits, start with Native Edit.

General

• Independent CREATE, Native Edit and Classic Img2Img architecture preserved.

• Improved execution isolation.

• Improved source-image handling.

• Improved Comfy Registry compatibility.

• Improved Missing Custom Nodes detection.

• Improved ComfyUI Manager installation and removal.

• Qwen3.5 Autoprompt installer included.

• Full Image Edit documentation included.

• Public release branding updated to v2.0.5.

VELVET VICE · KREA 2 VISION PROMPTER v2.0.5 is focused on making Krea 2 generation and image editing smarter, more predictable, easier to troubleshoot, easier to install and much cleaner to distribute.