A Personal Note

This is my first experiment with Qwen Image 2.1 Turbo, and I wanted to share my experience and the results of this hybrid INT8/W4A4 conversion.

To be perfectly honest, I'm not particularly blown away by Qwen Image 2.1 so far. It's certainly a capable model, but personally, I still prefer Krea 2 Turbo, which I find more visually appealing, with a more beautiful and pleasing overall aesthetic.

Of course, this is purely my personal opinion and a matter of taste. Everyone has their own preferences, and I respect the work behind Qwen Image 2.1.

Vince Fidelity is simply my first attempt to explore this architecture while preserving as much of the original INT8 model's visual quality as possible.

MODEL NAME: Qwen Image 2.1 – Vince Fidelity

VERSION: Turbo Hybrid INT8-W4A4 v1.0

TYPE: Checkpoint

Qwen Image 2.1 – Vince Fidelity is an experimental fidelity-focused hybrid quantization based on Qwen Image 2.1 Turbo INT8 ConvRot. The goal is to retain the visual character and quality of the original INT8 checkpoint while selectively reducing model size.

The transformer retains 176 INT8 matrices, with 16 matrices converted to packed 4-bit W4A4-style ConvRot-compatible representation (8 baseline selections plus 8 new selections). It is NOT a fully W4A4 checkpoint and is NOT a newly trained or fine-tuned model.

TESTED CONFIGURATION

- Frontend: Stable Diffusion Forge Neo

- Steps: 8

- CFG: 1.0

- Sampler: Euler

- Scheduler: Simple

- VAE: qwen_image_2.1_vae_bf16.safetensors (separate download)

- Text encoder: qwen3vl_8b_w4a8.safetensors (separate download)

- Tested GPU: NVIDIA RTX 3060 Laptop 6GB, using CPU offloading

- Negative prompt: ignored at CFG 1.0

VISUAL CHARACTER

The model was tested on natural skin textures, portraits, hands, detailed medieval scenes, wolf armor, fashion photography, mechanical macro subjects, gothic monochrome and nighttime neon lighting. Results looked strong in these samples. Visual similarity to the INT8 baseline is the target, not a guarantee of pixel-level equivalence.

COMPATIBILITY

Forge Neo was tested successfully. Other frontends, GPUs, text encoders, or LoRA combinations have not been validated. CPU offloading means a 6GB GPU is not sufficient to keep the entire pipeline in VRAM.

CREDITS AND RESTRICTIONS

Base: Qwen-Image-2.1-Turbo by the Qwen team.

Original: https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo

Conversion and testing: Vince AI.

Independent unofficial quantized derivative. Not endorsed by Qwen.

License: Qwen Research License Agreement. Non-commercial research/evaluation use only; commercial rights are not granted by this upload. Please read the license accompanying the release and the NOTICE file.

MODEL FILE SHA256

290e8a8a3c0b2ea8aa9eebb9442e315e23fb0ff381241303c56a8f5e5a0832b5

https://huggingface.co/Qwen/Qwen-Image-2.1/blob/main/LICENSE

Alternative Text Encoder — INT8 ConvRot

I also tested Vince Fidelity with the Qwen3VL 8B INT8 ConvRot text encoder, available from Comfy-Org, as an alternative to the Qwen3VL 8B W4A8 encoder.

The results were very encouraging, particularly for photorealistic portraits, fashion photography, natural skin textures, and clothing details.

In my tests, the INT8 ConvRot encoder produced convincing images with good facial consistency, realistic materials, and strong visual quality.

Both encoders remain usable, but the INT8 ConvRot version is worth trying if you prioritize image quality over memory efficiency.

Text encoders:

• INT8 ConvRot: Higher-precision alternative, visually promising in my tests.

• W4A8: Lighter option, recommended for limited-memory configurations.

Note: These are subjective visual observations, not a controlled benchmark proving that one encoder is consistently superior. https://huggingface.co/Comfy-Org/Qwen-Image-2.1/blob/main/text_encoders/qwen3vl_8b_int8_convrot.safetensors

Recommended Resolution Settings

Vince Fidelity has been successfully tested at 1024 × 1536, delivering excellent results for photorealistic portraits, retro pin-up photography, fashion editorials, and detailed full-body compositions.

Suggested Resolutions:

• 1024 × 1536 — Recommended for portraits, fashion, glamour, and full-body photography.

• 1024 × 1024 — Square portraits, close-ups, and product photography.

• 1152 × 1536 — Portrait compositions with more background detail.

• 1536 × 1024 — Landscapes, cinematic scenes, and environmental photography.

Recommended Generation Settings:

• Steps: 8

• CFG: 1.0

• Sampler: Euler

• Scheduler: Simple

• Hires.fix: Optional

• Upscaler: 4x-UltraSharp

Personal recommendation: 1024 × 1536 is my preferred resolution for Vince Fidelity, offering excellent visual quality and detailed vertical compositions.

Other resolutions are suggested starting points and may require additional testing depending on your hardware.