Hugging Face's logo

wangfuyun
/
consolver

Text-to-Image
Diffusers
diffusion-models
stable-diffusion
image-generation
fast-sampling
Model card Files
xet
Community
consolver
File size: 3,335 Bytes 
ee650ee
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
 
---
license: apache-2.0
base_model: runwayml/stable-diffusion-v1-5
tags:
- text-to-image
- diffusion-models
- stable-diffusion
- diffusers
- image-generation
- fast-sampling
library_name: diffusers
pipeline_tag: text-to-image
---
# Image Diffusion Preview with Consistency Solver (Google DeepMind)

[paper](https://arxiv.org/abs/2512.13592)  [code](https://github.com/G-U-N/consolver)   [huggingface](https://huggingface.co/papers/2512.13592)  [model](https://huggingface.co/wangfuyun/consolver)

# Quick Start

```python
Pythonimport torch
from diffusers import StableDiffusionPipeline, DDIMScheduler
from scheduler_ppo import PPOScheduler  # Provided in this repo
from huggingface_hub import hf_hub_download

# Download the trained factor_net checkpoint
factor_net_path = hf_hub_download(
    repo_id="wangfuyun/consolver",
    filename="model.ckpt"
)

model_id = "runwayml/stable-diffusion-v1-5"
prompt = "an astronaut riding a horse on the moon, highly detailed, 8k"
num_inference_steps = 8
guidance_scale = 3.0
seed = 43
height = width = 512

def load_pipeline(scheduler_type="ddim"):
    if scheduler_type == "ppo":
        scheduler = PPOScheduler(
            beta_end=0.012,
            beta_schedule="scaled_linear",
            beta_start=0.00085,
            num_train_timesteps=1000,
            steps_offset=1,
            timestep_spacing="trailing",
            order_dim=4,
            scaler_dim=0,
            use_conv=False,
            factor_net_kwargs=dict(embedding_dim=64, hidden_dim=256, num_actions=11),
        )
    else:
        scheduler = DDIMScheduler.from_pretrained(model_id, subfolder="scheduler", timestep_spacing="trailing")

    pipe = StableDiffusionPipeline.from_pretrained(
        model_id,
        scheduler=scheduler,
        safety_checker=None,
        # torch_dtype=torch.float16,  # Uncomment for GPU memory savings
    ).to("cuda")

    if scheduler_type == "ppo" and factor_net_path:
        weight = torch.load(factor_net_path, map_location="cpu")
        pipe.scheduler.factor_net.load_state_dict(weight)
        pipe.scheduler.factor_net.to("cuda")

    return pipe

generator = torch.Generator("cuda").manual_seed(seed)

# DDIM baseline (8 steps)
pipe_ddim = load_pipeline("ddim")
image_ddim = pipe_ddim(prompt, num_inference_steps=num_inference_steps, guidance_scale=guidance_scale,
                       generator=generator, height=height, width=width).images[0]
image_ddim.save("ddim_result.jpg")

# ConSolver (8 steps)
pipe_consolver = load_pipeline("ppo")
image_consolver = pipe_consolver(prompt, num_inference_steps=num_inference_steps, guidance_scale=guidance_scale,
                                 generator=generator, height=height, width=width).images[0]
image_consolver.save("consolver_result.jpg")
```




<div align="center">
  <table>
    <tr>
      <td align="center">
        <img src="https://github.com/user-attachments/assets/35f5f99a-ca5f-4919-82cf-04a67a2dbe13" alt="DDIM" width="80%" />
      </td>
      <td align="center">
        <img src="https://github.com/user-attachments/assets/6428a663-b488-4ecc-b79c-4fcb431d5630" alt="Consistency Solver" width="80%" />
      </td>
    </tr>
    <tr>
      <td align="center">
        <em>DDIM</em>
      </td>
      <td align="center">
        <em>ConsistencySolver</em>
      </td>
    </tr>
  </table>
</div>