README.md

metadata

title: Pornmaster Krea2
emoji: 🏆
colorFrom: indigo
colorTo: yellow
sdk: gradio
sdk_version: 6.22.0
python_version: '3.12'
app_file: app.py
pinned: false
short_description: Krea 2 PornMaster finetune space.

Pornmaster Krea 2

Krea 2 Turbo text-to-image on Gradio, running the ComfyUI backend directly (no ComfyUI server, no custom nodes). Deploy on ZeroGPU hardware.

Workflow: image_krea2_turbo_t2i.json Pattern: Run ComfyUI workflows for free with Gradio on Hugging Face Spaces

How it works

• app.py clones the ComfyUI backend (once), downloads the models, loads them at module scope, and runs the flattened workflow via direct node calls inside generate_image(), decorated with @spaces.GPU.
• UNet (diffusion model) comes from CivitAI: PornMaster-Krea2 (model 2735032, version V2.5 Turbo / 3171380, fp8). Companion models (text encoder, VAE, LoRA) come from Comfy-Org/Krea-2, which is not gated.

CivitAI version ladder

V2.5 (3171380) is currently Early Access on CivitAI (needs Buzz to unlock), so the app automatically tries, in order:

1. CIVIT_MODEL_VERSION (default 3171380, V2.5 Turbo fp8)
2. CIVIT_FALLBACK_VERSION (default 3112108, Turbo V2 FP8) - downloads fine today
3. Stock krea2_turbo_fp8_scaled.safetensors from Comfy-Org (last resort)

Once V2.5 is unlocked on the CivitAI account, it is used automatically (no code change needed). Set CIVIT_MODEL_VERSION=3112108 on the Space to force V2.

Setup

1. Create a Space with Gradio SDK (already configured in this repo).
2. Set hardware to ZeroGPU (large is fine). ZeroGPU PRO or grant required.
3. Add the CivitAI API key as a Space secret named CIVIT_API_KEY.
4. First cold start downloads ~17.4 GB of models (5-10 min) and clones ComfyUI. After that, ZeroGPU keeps the container disk, so resumes are fast.

Notes / tuning

• duration: @spaces.GPU(duration=120). Measure worst-case wall time and set duration = measured * 1.4. LLM prompt enhancement adds 20-60 s.
• ZeroGPU sizing: size="large" is the default (48 GB VRAM slice on Blackwell). Only switch to xlarge (2x quota) if it OOMs.
• Return values: numpy arrays are returned (never CUDA tensors), required across the ZeroGPU fork boundary.
• No HF token needed: none of the model repos are gated.

Why ComfyUI and not diffusers

Diffusers does have a Krea2Pipeline, but the official diffusers-format repos (krea/Krea-2-Turbo) are gated and 26 GB bf16, and there is no tooling to load raw CivitAI state dicts (single-file checkpoints) with diffusers for Krea 2. The ComfyUI backend loads the CivitAI .safetensors directly via UNETLoader with automatic architecture detection, and Krea 2 LoRAs work natively.