4CGT — 四彫画匠

Shichō Gashō — the art of sculpted mastery.

Anime image generation with DoRA anatomical correction, prompt fidelity up to 600 tokens, and deterministic output — same seed, same PNG hash across macOS, Windows, and Linux.

_{All generated with 4CGT. DoRA v0.271. 42 steps. Same seed = same image on any machine.}

---

SDXL v-prediction frontend built around NoobAI XL V-Pred 1.0. DoRA weight-decomposed adapters for anatomical accuracy. BF16/FP32 lossless pipeline — FP16 is rejected on load. Deterministic output: same seed produces the same PNG hash across macOS, Windows, and Linux. Long prompt support up to 600 tokens (8 chunks × 75 tokens) with A1111 weight syntax. Integrated 2D-to-3D Gaussian Splat conversion via Apple Sharp. OpenPose ControlNet for v-prediction at conditioning scale 2.0. Runs locally on 6 GB VRAM (GTX 1060) through 32 GB (RTX 5090) and Apple Silicon M1–M4.

Quick start

Requires Python 3.11 or 3.12 (not 3.13) and PyTorch >= 2.7.0.

git clone https://github.com/teenu/4cgt.git && cd 4cgt

# Install PyTorch (pick your GPU — see full list below)
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124

# Install dependencies
pip install -r requirements.txt

# Download assets (see tables below for filenames, hashes, and sources)
# Place them: model in repo root, adapters in dora/, CSVs in style/

# Launch
python main.py

Opens a Gradio web UI at http://localhost:7860.

CLI mode

# DoRA None mode auto-applies: 42 steps, CFG 5.5261, rescale 0.6092, start step 3
python main.py --cli \
  --prompt "very awa, masterpiece, best quality, year 2024, newest, highres, absurdres, 1girl, silver hair, red eyes, black dress, night sky, stars, cinematic lighting" \
  --width 1216 --height 832 \
  --enable-dora --dora-adapter 0 --seed 7777

Assets

4CGT requires third-party model assets. Download from their original sources and verify SHA-256 hashes to ensure quality parity.

Base model — place in repo root

Filename	Size	SHA-256	Source
NoobAI-XL-Vpred-v1.0.safetensors	6.6 GB	ea349eea...e02819	Laxhar/noobai-XL-Vpred-1.0

DoRA adapters — place in dora/

Filename	Size	SHA-256	Source
noobai_vp10_stabilizer_v0.271_fp16.safetensors	43.7 MB	9567b54e...004fbc	Civitai: reakaakasky
noobai_vp10_stabilizer_v0.280a_fp16.safetensors	63.6 MB	99db2d0b...1e2318	Civitai: reakaakasky

Style CSVs — place in style/

Filename	Size	SHA-256	Source
danbooru_character_webui.csv	37.2 MB	4925522e...45ac1	Laxhar/noob-wiki
danbooru_artist_webui.csv	31.3 MB	a6dad048...ad741	Laxhar/noob-wiki
e621_character_webui.csv	27.9 MB	f18d33a3...85a91	Laxhar/noob-wiki
e621_artist_webui.csv	12.5 MB	87a78e28...6fa6	Laxhar/noob-wiki

ControlNet (optional) — place in controlnet/

Filename	Size	SHA-256	Source
openpose_fp32.safetensors	4.7 GB	9e763e0b...a50fc	xinsir/controlnet-openpose-sdxl-1.0

Full SHA-256 hashes

ea349eeae87ca8d25ba902c93810f7ca83e5c82f920edf12f273af004ae02819  NoobAI-XL-Vpred-v1.0.safetensors
9567b54e807c004eef3f89b268a8d61d76d0b62c0061484fc1816a03fc004fbc  noobai_vp10_stabilizer_v0.271_fp16.safetensors
99db2d0bf94c05777304bf08aa8d25c3ce8a99d1c1def4cbb81c06eddb1e2318  noobai_vp10_stabilizer_v0.280a_fp16.safetensors
4925522e2fddde5ed1815aa71b93a396fdd4b5c66c1345d394d01a6661c45ac1  danbooru_character_webui.csv
a6dad04843d2c3aaabeaba43fa95d33270e27ddd9d782ba099443598778ad741  danbooru_artist_webui.csv
f18d33a3ef47761d4991844d30f6c0826c4e471d21f8cc59e486eb34da785a91  e621_character_webui.csv
87a78e2811b4d679a8227759c10d61be63f7babaf49322ce7c780b88ef9d6fa6  e621_artist_webui.csv
9e763e0b0160050a3ade517d3efb80789bc8a0ba12c6a10b923b0eff242a50fc  openpose_fp32.safetensors

These assets carry their own licenses. See each source for terms.

How it works

DoRA anatomical stabilizers

Weight-Decomposed Low-Rank Adaptation adapters trained for v-prediction. Three modes:

• None mode — 42 steps, CFG 5.5261, rescale 0.6092, adapter strength 1.0, DoRA activates at step 3.
• Optimized — 34 steps, CFG 4.2, rescale 0.55, adapter strength 1.0. Binary schedule frozen: [0,0,0,0,0,0,1,1,1,0,1,1,0,0,0,0,0,1,1,0,0,0,1,0,0,0,0,1,1,1,1,1,1,1].
• Manual — per-step binary toggle grid. Each of the N steps gets a 0 (off) or 1 (on) value via interactive UI grid or CSV string.

Lossless precision pipeline

FP16 is rejected on load. Not downcast, not warned — rejected.

The pipeline is BF16 or FP32 only. VAE always decodes in FP32. The result: no precision-dependent artifacts, no platform-dependent drift. Same seed produces the same PNG file hash on macOS, Windows, and Linux.

Long prompts

Standard CLIP cuts off at 77 tokens. 4CGT extends this to 600 tokens (8 chunks × 75 usable tokens) via sd_embed, with full A1111-compatible syntax: (emphasis:1.2), [de-emphasis], ((nested weights)).

Image to 3D

Integrated Apple Sharp pipeline. Generate an image, convert it to a 3D Gaussian Splat (.ply), optionally render a camera flythrough video — all without leaving the UI. Sharp runs in a subprocess so it doesn't evict SDXL from VRAM.

ControlNet pose control

OpenPose skeleton conditioning optimized for v-prediction. Default conditioning scale is 2.0 (vs. 1.0 in standard pipelines) because v-pred models need stronger guidance.

GPU compatibility

GPU	BF16	Notes
RTX 5090 (Blackwell)	Native	CUDA 12.8+, PyTorch >= 2.7.0
RTX 30xx/40xx (Ampere/Ada)	Native
RTX 20xx (Turing)	Upcast to FP32	Same weights, FP32 arithmetic
Apple Silicon (M1–M4)	Native (AMX)
GTX 1060+ (6 GB VRAM)	Upcast to FP32	Auto CPU offload at < 8 GB VRAM

PyTorch install by GPU

# RTX 5090
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128
# RTX 30xx/40xx
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124
# RTX 20xx
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
# Apple Silicon
pip install torch torchvision torchaudio
# CPU only
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu

Benchmarks

GPU	Resolution	Steps	Time
RTX 5090 32GB	1216x832	42	~25s
RTX 3060 12GB	1216x832	35	~30s
RTX 2060 6GB	1216x832	35	~90s (CPU offload)

Troubleshooting

Windows transformers import error — corrupted pip cache. Fix: pip uninstall transformers -y && pip install transformers>=4.40.0

No model found — place NoobAI-XL-Vpred-v1.0.safetensors in the repo root or use --model-path

CUDA out of memory — reduce resolution to 1024x1024 or 832x1216, reduce steps to 25, or close other GPU apps. CPU offloading kicks in automatically when VRAM < 8 GB.

Windows path too long — enable long paths: New-ItemProperty -Path "HKLM:\SYSTEM\CurrentControlSet\Control\FileSystem" -Name "LongPathsEnabled" -Value 1 -PropertyType DWORD -Force then restart.

License

AGPL-3.0. Free to use, modify, and distribute. If you run a modified version as a network service, share your source.