metadata

license: apache-2.0
base_model: Qwen/Qwen-Image
language:
  - en
  - zh
pipeline_tag: text-to-image
library_name: diffusers
widget:
  - text: >-
      A close-up portrait of a dog with black, brown, and white fur, a white
      stripe on its forehead, and brown and black markings on its ears, is
      looking directly at the camera with a serious expression. The dog has
      brown eyes with black pupils and a black nose, and its ears are large and
      pointed. The background is blurred and appears to be an outdoor setting
      with green and brown grass and a light grey sky.
    output:
      url: examples/Qwen-Image_00133_.png
  - text: >-
      Close-up food photo of a hybrid snail composed entirely of glossy sticky
      cinnamon buns. The shell is made from a puffy perfectly swirled cinnamon
      bun covered in a thick glossy white glaze. Baked edges with a jagged
      cinnamon bun texture slightly caramelized, dark cinnamon filling inside,
      rich golden brown color. The glaze drips down in thick sweet drops, the
      snail tendrils are made of twisted cinnamon dough, glistening with icing
      sugar, the glaze reflects warm, natural light. The scene is shot in a
      soft, fuzzy kitchen setting, with a hint of freshly baked pastries in the
      background.
    output:
      url: examples/Qwen-Image_00134_.png
  - text: >-
      8-bit pixel art of a pidgeon wearing a lab coat, and a tie. The background
      is large computer server room. The lighting is dark, with most light
      hitting the servers and not the pidgeon.
    output:
      url: examples/Qwen-Image_00136_.png
  - text: >-
      A long tunnel with a high ceiling is seen dimly lit, illuminated by a
      single fluorescent light fixture at the end of the tunnel. The tunnel
      walls are made of corrugated metal and are lined with copper pipes. On the
      left wall, there is a yellow warning sign with a black exclamation mark
      and the text "WARNING - MILITARY TESTING" in black letters. To the right
      of the warning sign, on the right wall, is a green control panel with
      various knobs and switches, and a black and yellow warning tape is
      attached to the control panel. The floor is dark and wet, reflecting the
      light from the fluorescent light. A metal grate is visible on the floor.
    output:
      url: examples/Qwen-Image_00137_.png
tags:
  - Qwen-Image
  - distillation
  - LoRA
  - merge

50/50 merge of the 4-step and 8-step LoRA

Prompt
A close-up portrait of a dog with black, brown, and white fur, a white stripe on its forehead, and brown and black markings on its ears, is looking directly at the camera with a serious expression. The dog has brown eyes with black pupils and a black nose, and its ears are large and pointed. The background is blurred and appears to be an outdoor setting with green and brown grass and a light grey sky.

Prompt
Close-up food photo of a hybrid snail composed entirely of glossy sticky cinnamon buns. The shell is made from a puffy perfectly swirled cinnamon bun covered in a thick glossy white glaze. Baked edges with a jagged cinnamon bun texture slightly caramelized, dark cinnamon filling inside, rich golden brown color. The glaze drips down in thick sweet drops, the snail tendrils are made of twisted cinnamon dough, glistening with icing sugar, the glaze reflects warm, natural light. The scene is shot in a soft, fuzzy kitchen setting, with a hint of freshly baked pastries in the background.

Prompt
8-bit pixel art of a pidgeon wearing a lab coat, and a tie. The background is large computer server room. The lighting is dark, with most light hitting the servers and not the pidgeon.

Prompt
A long tunnel with a high ceiling is seen dimly lit, illuminated by a single fluorescent light fixture at the end of the tunnel. The tunnel walls are made of corrugated metal and are lined with copper pipes. On the left wall, there is a yellow warning sign with a black exclamation mark and the text "WARNING - MILITARY TESTING" in black letters. To the right of the warning sign, on the right wall, is a green control panel with various knobs and switches, and a black and yellow warning tape is attached to the control panel. The floor is dark and wet, reflecting the light from the fluorescent light. A metal grate is visible on the floor.

My recommended settings

LoRA Strength: 0.9 (or possibly even lower)
Steps: 16
Sampler: DEIS
Scheduler: KL Optimal
Shift: None (I removed the node, since it made no difference after I swapped to KL Optimal scheduler.)

Reason for making

The 4-step LoRA does fairly well at 4 steps, but it cannot go higher than 4 steps without overcooking the image, and even at 4 steps the image feels a little cooked.

4-step lora | 4 steps vs. 8 steps vs. 16 steps | [1536x1536, no shift, lora strength 1, deis, kl_optimal, seed 187]

The 8-step LoRA on the other hand is very undercooked at 4 steps, still a little undercooked at 8 steps, but handles higher step counts like 16 really well, but feels a little undercooked overall.

8-step lora | 4 steps vs. 8 steps vs. 16 steps | [1536x1536, no shift, lora strength 1, deis, kl_optimal, seed 187]

Merging these two together results in being able to do 16 without overcooking or undercooking. It feels just about right, especially if you load at 90% strength.

merged lora | 4 steps vs. 8 steps vs. 16 steps | [1536x1536, no shift, lora strength 1, deis, kl_optimal, seed 187]