Skip to content
Images

ovb-b-17500-best-full

standalone Qwen-Image-2.1 pipeline with baked-in LoRA

2 min readupdated

A ready-to-use QwenImage21Pipeline (7.12B parameters) for image-to-image editing and analysis. It takes an RGB image and a text instruction to output a fixed-palette depth map, perform semantic or instance segmentation, or generate a new image. Released by gatilin, it merges the Melo111/Fusion-C1-KQ0r128-60-40 base model with the ovb-i108-b LoRA adapter (a lightweight add-on trained on top of a base model) at training step 17,500. Unlike the standalone adapter, the weights are baked directly into the main transformer, so you don't need to download the base model or load the adapter separately at runtime.

Key features

  • Fully self-contained: includes the merged bfloat16 transformer, frozen text encoder, VAE, tokenizer, and scheduler.
  • Clean weight merge: fused via PEFT's merge_and_unload(safe_merge=True) across 224 modules (including attention layers to_q, to_k, to_v, and img_mlp) from an original rank-64 rsLoRA.
  • Multi-task support: handles depth map visualization, segmentation (semantic, instance, and referring/prompt-guided), and prompt-based image generation.
  • Fine-tuning ready: the transformer component can be loaded on its own as a clean initialization checkpoint for a new training run.

What you can do

  • Generate depth maps and segmentation masks: feed RGB photos with open-vision-banana project prompts to produce visual predictions.
  • Run locally with minimal dependencies: load the pipeline in diffusers scripts with CPU offloading (enable_model_cpu_offload) without fetching parent repositories.
  • Use as a fine-tuning base: extract the transformer module and start fresh training runs without leftover optimizer states.

Example

import torch
from diffusers import QwenImage21Pipeline

repo = "gatilin/ovb-b-17500-best-full"
pipe = QwenImage21Pipeline.from_pretrained(repo, torch_dtype=torch.bfloat16)
pipe.enable_model_cpu_offload()
Read moreDetails
Export verification: checked SHA-256 hashes for all files, tensor loading accuracy, and LoRA merge math across every module.
Limitations: full offline pipeline execution and image generation were not tested by the author after merging (image_inference: false), and benchmark scores were not re-evaluated.
License: governed by the Qwen Research License (for research and evaluation only; non-commercial use).
Discuss Telegram 0 17

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 13 GB · ready files16 GB
  • FP8 / INT8recommendedweights 6.6 GB8.8 GB
  • 4-bitweights 3.7 GB5.6 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 3080 · BF16 — from RTX 4090

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

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.

    # gatilin/ovb-b-17500-best-full
    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 gatilin/ovb-b-17500-best-full --local-dir ~/models/ovb-b-17500-best-full
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to ovb-b-17500-best-full
  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.

  • SageAttention 3v2.2.0from October 28, 2025

    even faster on RTX 50 / Blackwell

    FP4 attention for Blackwell. Faster than SageAttention 2, but the authors suggest v2 when accuracy matters.

    How to enable: Build from the sageattention3_blackwell folder of the repo; in ComfyUI via KJNodes.

    Earlier generations (2)
    • SageAttention 2++ — Quantised (INT8/FP8) attention with no visible quality loss. The main speed-up for image and video generation.
    • SageAttention 1 — First version: INT8 for Q·K. Still works, but 2++ is faster and more accurate.
  • 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 (3)
    • 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.
    • xFormers — The older way to speed up attention, works almost everywhere. Sage or Flash are faster where available.
  • cache-dit (DBCache)v1.5.6from November 20, 2026new release

    1.5–3× at nearly the same quality

    Reuses block outputs between diffusion steps. Supports most DiT models.

    How to enable: pip install cache-dit; one line cache_dit.enable_cache(pipe) for diffusers.

    Earlier generations (2)
    • MagCache — Skips steps by magnitude of change. More accurate than TeaCache, no calibration.
    • TeaCache — The first popular diffusion step cache. Works, newer methods are more accurate.
  • Nunchaku (SVDQuant 4-bit)v1.2.1from January 25, 2026

    ≈3× faster and 3.5× less memory

    4-bit versions of popular image models with almost no quality loss (FP4 on RTX 50).

    How to enable: Ready weights for supported models; in ComfyUI use ComfyUI-nunchaku.

  • 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

  1. Qwen-Image-2.1family root
  2. Fusion-C1-KQ0r128-60-40
  3. ovb-b-17500-best-full

New posts about versions of Qwen-Image-2.1 (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 4All on Hugging Face

About the model

gatilin/ovb-b-17500-best-full
Author
gatilin
Category
Images
Released
Posted

On NeyroLearn

Hugging Face

170

Ask AI about this model

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

Comments

Images

DeFLOCNet

DeFLOCNet
Images

Pixie P32

Pixie P32