Instructions to use e-n-v-y/Qwen-Image-2.1-Fix-v2.0 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use e-n-v-y/Qwen-Image-2.1-Fix-v2.0 with Diffusers:

pip install -U diffusers transformers accelerate

import torch
from diffusers import DiffusionPipeline

# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("Qwen/Qwen-Image-2.1", dtype=torch.bfloat16, device_map="cuda")
pipe.load_lora_weights("e-n-v-y/Qwen-Image-2.1-Fix-v2.0")

prompt = "A dramatic, high-fantasy digital painting depicting a colossal, ferocious owlbear in full predatory stance, its massive, feathered wings partially unfurled as if preparing to strike or take flight. The creature’s body merges the muscular bulk of a bear with the sharp, predatory features of an owl — a leathery, scaly hide mottled with dark browns and grays, feather tufts erupting from its shoulders and back, and piercing, glowing amber eyes that burn with primal rage. Its taloned forepaws are raised, claws gleaming with menace, while its beak-like snout drips with saliva and reveals dagger-like teeth. The scene is set in a twilight forest clearing, with twisted ancient trees casting long, ominous shadows, and mist curling through the underbrush, enhancing the eerie, mythic atmosphere.  The lighting is cinematic and moody, with a low-angle sun casting golden rays through the canopy, illuminating the owlbear’s fur and feathers while leaving the surrounding forest in deep, velvety shadow. Behind the creature, a distant, crumbling stone tower looms through the mist, hinting at a forgotten realm or cursed fortress. The ground is littered with broken branches, scattered bones, and glowing mushrooms that pulse with bioluminescence, adding a touch of magical realism. The owlbear’s fur and feathers are rendered with intricate detail — individual strands catching the light, with textures ranging from coarse, bear-like hide to soft, downy feathers — creating a tactile, lifelike surface that contrasts with the surreal, otherworldly setting.  The composition is dynamic and immersive, with the owlbear positioned slightly off-center to draw the viewer’s eye into the frame, its gaze fixed directly forward as if challenging the viewer. The brushwork is painterly yet digitally precise, blending realistic anatomy with fantastical exaggeration — the creature’s wings are oversized and slightly tattered, its claws are unnaturally long and sharp, and its eyes radiate an almost supernatural intensity. The palette is rich and saturated, dominated by deep forest greens, charcoal grays, and fiery amber highlights, with subtle iridescent glints on the feathers and mist. The overall mood is one of awe, danger, and ancient magic — a mythic beast emerging from the shadows to command reverence and fear.,"
image = pipe(prompt).images[0]

Notebooks
Google Colab
Kaggle
Local Apps Settings
Draw Things
DiffusionBee

Demo for this model on Spaces

by multimodalart - opened 5 days ago

Discussion

multimodalart

5 days ago

Hi @e-n-v-y 🤗

I'm Apolinario, from the open-source team at Hugging Face. Congrats and thanks for open-sourcing e-n-v-y/Qwen-Image-2.1-Fix-v2.0 on the Hub! We were excited about this work and built with an agent an interactive demo app of it on Hugging Face Spaces, running on a free ZeroGPU infrastructure.

Here's a link to the demo: https://huggingface.co/spaces/hugging-apps/qwen-image-21-fix-lora

We would love to transfer this demo to you or your organization. Would you like this demo to live under your own account or organization? If so just let me know here which username to transfer to, and we'll transfer the Space over to you, we hope it can give your work more visibility, discoverability and allows folks to try it out.

(If you have any questions or just want to chat more about this, you can find me on Twitter, LinkedIn or apolinario @ huggingface.co)

Cheers,

Poli

e-n-v-y

Owner about 8 hours ago

Hi!

Thanks for doing this! I'm noticing that for some reason the generations are coming out really dark, or really light.

multimodalart

about 6 hours ago

•

edited about 6 hours ago

Thanks for flagging this, @e-n-v-y ! You were right, the exposure was off. It's fixed now: https://huggingface.co/spaces/hugging-apps/qwen-image-21-fix-lora

What was wrong: The LoRA itself was loaded correctly (same alpha/rank scaling and gate/up split as ComfyUI, bf16). The problem was in how the demo ran CFG 3.5 with your negative prompt. Diffusers' Qwen-Image-2.1 pipeline does plain true-CFG with no renormalization. Combined with only 20 Euler steps, the guided prediction overshot and whole images drifted very dark or blown-out light. This also happened on the base model without the LoRA.

Fix:

The CFG output is now rescaled per token to the conditional prediction's norm, the same recipe the original Qwen-Image pipeline uses.
The default is now 40 Euler steps, the same number of model evaluations as your workflow's 2nd-order res_2m sampler at 20 steps.

Your card prompts now come out with exposure close to your comparison images. Let me know if anything still looks off!

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment