Skip to content
Computer vision

Artsy RLHF

A LoRA for drawing via code generation

2 min readupdated

The final checkpoint of the Artsy project is a LoRA adapter (a lightweight add-on on top of a base model) built on Qwen/Qwen3.5-4B by Rashik24. Instead of generating pixels directly, the model writes Python code and calls drawing environment tools, while an external environment handles rendering with preview support. Trained across three stages (SFT → RLVR → RLHF using rewards from the HPSv3 human preference model), the adapter significantly outperforms the base Qwen3.5-4B in reliability and the visual quality of the final illustrations. It is a great fit for agentic system developers, multimodal LLM researchers, and anyone experimenting with procedural graphics generation through code.

Key features

  • Code-based graphics generation: The model outputs Python scripts and tool calls, and the final image is produced when they execute in a sandbox.
  • Optimized for render quality: A final RLHF stage across 1,000 prompts scored with HPSv3 boosts aesthetics without degrading coding ability.
  • Iterative workflow: Retains full support for generation, visual previewing, and final result submission.
  • Updated output-head architecture: Must be loaded via the PEFT library while keeping the base model's separate input embeddings intact.

What you can do

  • Integrate into Artsy and ArtCanvas: Use it as your main checkpoint for creating illustrations from complex text prompts.
  • Embed into an agentic pipeline (mini-SWE): Run model-generated code inside an isolated sandbox with a preview-based self-verification loop.
  • Compare SFT, RLVR, and RLHF training stages: Test how preference optimization (PPO + HPSv3) affects the valid render rate.

Example

from transformers import AutoProcessor, Qwen3_5ForConditionalGeneration
from peft import PeftModel

base_id = "Qwen/Qwen3.5-4B"
rev = "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a"
processor = AutoProcessor.from_pretrained(base_id, revision=rev)
base = Qwen3_5ForConditionalGeneration.from_pretrained(base_id, revision=rev, dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(base, "Rashik24/rlhf-experiment10-genai1000-linear-epoch1-kl-budget-recovery").eval()
Read moreDetails
Benchmark (150 complex GenAI-Bench prompts, 12-step limit): 149/150 valid images and an average HPSv3 score of 0.539, compared to 102/150 and 0.094 for the base Qwen3.5-4B.
Training: Rank-128 LoRA, 32 updates (1 epoch) on 1,000 prompts, batch size 32, 4 trajectories per prompt, lr 1e-5, KL penalty 0.01.
Limitations: Requires Artsy instructions, the ArtCanvas API, and a sandbox; you must pass enable_thinking=False in the chat template.
Compatibility: Tested with Transformers 5.5.4 and PEFT 0.21.3.
Discuss Telegram 0 23

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 9.8 GB · ready files11 GB
  • FP8 / INT8recommendedweights 5.4 GB6.5 GB
  • 4-bitweights 3.5 GB4.4 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 5060 Ti 8GB · BF16 — from RTX 5070

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

This is a LoRA: counted together with its base model Qwen3.5-4B.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # Rashik24/rlhf-experiment10-genai1000-linear-epoch1-kl-budget-recovery
    # ===== Tokens: paste yours between the quotes =====
    export HF_TOKEN=""        # huggingface.co/settings/tokens, for gated models
    # ==========================================
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download Qwen/Qwen3.5-4B --local-dir ~/models/Qwen3.5-4B
    hf download Rashik24/rlhf-experiment10-genai1000-linear-epoch1-kl-budget-recovery --local-dir ~/models/rlhf-experiment10-genai1000-linear-epoch1-kl-budget-recovery
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to Qwen3.5-4B
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

Requires the base model

Qwen/Qwen3.5-4B

This model is a LoRA / adapter. Download the base model first, then load this one on top.

  1. Qwen3.5-4B-Basefamily root
  2. Qwen3.5-4B
  3. rlhf-experiment10-genai1000-linear-epoch1-kl-budget-recovery

New posts about versions of Qwen3.5-4B-Base (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 12All on Hugging Face
Merges 1All on Hugging Face

About the model

Rashik24/rlhf-experiment10-genai1000-linear-epoch1-kl-budget-recovery
Author
Rashik24
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

230

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments