sekkit/UltraTex-Qwen-Image-2.1

UltraTex Qwen-Image 2.1 port

This is a port of UltraTex (MIT, Yibo Zhang) to the Qwen-Image 2.1 backbone, trained by Atlas 3D (orwelian84). The released UltraTex weights only cover FLUX.1-dev and FLUX.2-Klein, so this is, to our knowledge, the first Qwen-Image multi-view texture LoRA.

Code: github.com/Atlas3DSS/UltraTex (fork of upstream with the Qwen-Image 2.1 and Krea 2 backbones on the default branch). The code/ folder here is a snapshot of the same files.

The method is UltraTex's. This release is the backbone port (code/) plus the trained weights (qwen21/) and samples.

Files

Path	Size	Description
qwen21/lora/dit_lora.safetensors	608 MiB	Multi-view texture LoRA, checkpoint-15000
qwen21/lora/train_args.json	1.5 KiB	Training arguments for the LoRA
qwen21/decoder_step1200.safetensors	988 MiB	Foreground-aware RGBA decoder, step 1200
qwen21/decoder_train_log.jsonl	89 KiB	Decoder training log
qwen21/text_cond.safetensors	64 KiB	Cached empty-prompt text conditioning
qwen21/text_cond.json	122 B	Metadata for the text conditioning cache
code/	~0.2 MiB	Qwen-Image 2.1 backbone and pipeline changes, scripts, tools, tests (apply on top of upstream UltraTex)
code/README.md	15 KiB	UltraTex README as of this port
samples/eval_grid_ckpt15000.webp	495 KiB	Evaluation grid, checkpoint-15000
samples/albedo_canvas_ckpt15000.webp	651 KiB	Albedo canvas output, checkpoint-15000
samples/fox_flux2_left_vs_qwen21_right.glb	21 MiB	Fox comparison: stock FLUX.2-Klein UltraTex (left) vs this LoRA (right)
samples/qwen21_ckpt15000_side_by_side.glb	101 MiB	Side-by-side GLB, this LoRA at checkpoint-15000
samples/fox_adventurer_v2_skinned.glb	12 MiB	Skinned sample mesh
samples/axolotl_suit_skinned.glb	14 MiB	Skinned sample mesh
samples/robot_companion_skinned.glb	12 MiB	Skinned sample mesh

Training

• LoRA rank 64, alpha 64
• 15,000 steps, 1024 px per view, batch 1
• lr 5e-5, constant; bf16; seed 42
• Sparse attention top-k 0.2
• Trained on G-buffer TexVerse (YiboZhang2001/G-buffer-TexVerse, gated)
• Foreground-aware RGBA decoder fine-tuned for 1,200 steps; foreground PSNR 38.65 dB
• Text conditioning is the cached empty prompt from Qwen3-VL-8B

Usage

1. Clone Atlas3DSS/UltraTex (or apply code/ on top of upstream UltraTex).
2. Put qwen21/ at checkpoints/qwen21/.
3. Set the paths in scripts/mv_env.sh to a ComfyUI checkout with Qwen-Image 2.1 support.
4. Run multi-view inference:

BACKBONE=qwen21 LORA=checkpoints/qwen21/lora DECODER_CKPT=checkpoints/qwen21/decoder_step1200.safetensors bash scripts/inference_mv.sh

Or run one mesh end to end:

BACKEND=qwen21 bash scripts/skin_mesh.sh MESH.glb REF.png OUT_DIR

Base weights are not included: Qwen-Image 2.1 (bf16), its VAE, and Qwen3-VL-8B must be obtained separately.

Samples

The GLBs are TRELLIS meshes generated from Krea 2 concept art, textured with this LoRA. The fox comparison GLB has stock FLUX.2-Klein UltraTex on the left and this LoRA on the right.

Credits

Method: UltraTex, by Yibo Zhang and contributors (MIT), github.com/yiboz2001/UltraTex. Qwen-Image 2.1 port, training, and weights: Atlas 3D.