a coffee is appreciated: ☕ Support me on Ko-fi.
Downloads
| Version | File size | Download |
|---|---|---|
| Original FP8 | 13.14 GB | Krea2Realism_V1.safetensors |
| W4A8 | 7.93 GB | Krea2Realism_V1_W4A8.safetensors |
| INT8 ConvRot | 13.32 GB | Krea2Realism_V1_INT8_ConvRot.safetensors |
W4A8 version
A smaller version of this checkpoint is now available: 4-bit weights with 8-bit activations, about 40% smaller than the original FP8 file. It was converted from the FP8 release, with sensitive layers retained at higher precision.
Place it in ComfyUI/models/diffusion_models and use the standard Load Diffusion Model node with weight_dtype set to default. Use an updated ComfyUI/comfy-kitchen installation that supports asym_w4a8_int8; no custom loader is required. Keep your usual Krea 2 text encoder and VAE.
Tested settings: Euler + Simple, 6 steps, CFG 1.0, 1152 × 1792 (about 2MP). Two matching prompts and seeds were compared with the FP8 version on an RTX 5090. Results can differ in pose and fine detail. RTX 5060 performance has not been tested directly, and file size is not total VRAM usage—offloading may still be needed.
FP8 vs. W4A8 comparison
Each pair shows FP8 on the left and W4A8 on the right, with matching prompts, seeds, and settings. Face crops are shown below.
INT8 ConvRot version
INT8 ConvRot is now available (13.32 GB). This is a separate conversion from the original FP8 release, with sensitive layers retained at higher precision. It is slightly larger than the FP8 file; W4A8 remains the smaller download.
Place it in ComfyUI/models/diffusion_models and use the standard Load Diffusion Model node with weight_dtype set to default. Use an updated ComfyUI/comfy-kitchen installation with int8_tensorwise ConvRot support. Keep your usual Krea 2 text encoder and VAE; no custom loader is required.
Tested settings: Euler + Simple, 6 steps, CFG 1.0, 1152 × 1792 (about 2MP). Standard loading and two matched-prompt generation tests passed on an RTX 5090. Speed and VRAM usage on other GPUs have not been measured.
FP8 vs. INT8 ConvRot comparison
Each pair shows FP8 on the left and INT8 ConvRot on the right, using identical prompts, seeds, and settings. Matching detail crops appear below. These examples look close, with small changes in hand position, facial detail, and background objects.
Showcase
Model Release Notes
This model is a merge of a custom Realism Checkpoint and V3 LoRA.
Because it was trained on high-resolution, high-quality images, it performs best at 2MP resolution. Generating at 1MP may introduce minor noise, though this is easily fixed using RTX upscaling nodes or similar post-processing tools.
⚠️ Note: This is a full NSFW checkpoint. Be sure to describe clothing clearly and specifically in your prompt—otherwise, expect some surprises! :D
Key Improvements & Features
- Full NSFW Support: Offers complete NSFW capabilities while maintaining the quality standards of V2 and V3.
- Natural Skin Tones & Textures: Unlike many other NSFW checkpoints, this model avoids synthetic "plastic" rendering and supports a diverse range of skin tones without locking onto a single default look.
Prompting Guidelines
- Style: Operates similarly to standard Krea 2 models. Overall image quality depends heavily on your prompting style.
- Format: Natural language prompts work best. Short or single-sentence prompts work fine, but they leave more interpretation up to the model, so give it enough detail if you have a specific vision in mind!
Recommended Settings
- Sampler / Scheduler:
Euler+Simple - Steps: 4–8 steps (6 steps is the sweet spot). You can go higher if you prefer, but 6 is generally all you need.
- Resolution: 2MP recommended.
What’s Next?
I am currently putting together a full ComfyUI workflow (WF) along with a dedicated custom node for prompting techniques that I’ve been developing. I will share and upload those soon—hang tight!
For now, simply drop this model into your existing pipeline as a swap-in replacement for your current checkpoint.
Give it a try and share your generated images! Enjoy!