Comments (8)
Does anyone have an example of how to use the iterative inpainting function?
from consistency_models.
@arittenbach I tried the following based on your recommendation but was not able to get the same as Figure 10 (mainly just got noise within the inpainted area.
Have you had any updates?
@aarontan-git Alright after playing around for awhile I think I figured it out. At the very least it generates results similar to what is seen in figure 10 in the paper using the LSUN bedroom LPIPS model
using the iterative inpaint function in cm/karras_diffusion.py :
def iterative_inpainting(
distiller, -> should be same model given to sample function in karras_sample function when generating imagers (variable denoiser)
images, -> stack of images to inpaint, tensor with (batch_size, 3, image_size, image_size), should be in range -1 to 1, i used the exact output of the karras_sample function
x, -> noise generated by generator, same size as image (batch_size, 3, image_size,image_size)
ts, -> from algorithm 4 in appendix of paper, ts should be range(2,41) if steps is 40
t_min=0.002, (i left this as is)
t_max=80.0, (i left this as is)
rho=7.0, (i left this as is)
steps=40, (i left this as is)
generator=None, (same generator used to generate images in karras_sample function)
):
otherwise, i used same parameters as in launch script for generating images using lsun_bedroom lpips with multistep.
hope this is clear
from consistency_models.
Does anyone have an example of how to use the iterative inpainting function?
Also looking for the same - @arittenbach let me know if you've had any insight?
from consistency_models.
@jS5t3r In case you're still looking into this or for anyone else wondering, you should not pass distiller=model
in iterative_inpainting.
distiller
should be the denoiser
function in karras_sample
, something like this:
def denoiser(x_t, sigma):
_, denoised = diffusion.denoise(model, x_t, sigma, **model_kwargs)
if args.clip_denoised:
denoised = denoised.clamp(-1, 1)
return denoised
sample, original = iterative_inpainting(
distiller=denoiser,
images=ref_img,
x=generator.randn(*ref_img.shape, device=dist_util.dev()),
ts=ts,
t_min=0.002,
t_max=80.0,
rho=diffusion.rho,
steps=args.steps,
generator=generator
)
After doing this I get inpainting samples that resemble Fig 10 of the paper! Basically in your code, you pass in model
directly, thus bypassing the skip connections (Eq. 5 of paper) that defines the Consistency Model that is defined in diffusion.denoise
, which is incorrect and leads to artifacts.
from consistency_models.
Is it planned to release the code for the Zero-Shot Image Editing? In particular, I'd be interested in the super-resolution.
Best regards
The function iterative_colorization/inpainting/superres are contained in karras_diffusion.py, I think maybe this is what you are looking for.
from consistency_models.
thanks for your code in karras_diffusion.py. But I've noticed that the effect of the superres heavily rely on some hyperparameters when using cd_bedroom256_l2.pt, say ts in https://github.com/openai/consistency_models/blob/6d26080c58244555c031dbc63080c0961af74200/cm/karras_diffusion.py#LL838C8-L838C8
So are there some recommended hyperparameter for superres? or a scripts like launsh.sh?
from consistency_models.
Best wishes.
from consistency_models.
Launch File:
mpiexec -n 1 python image_sample_inpainting.py --batch_size 32 \
--training_mode consistency_distillation \
--sampler multistep \
--ts 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 \
--steps 40 \
--model_path checkpoints/cd_bedroom256_lpips.pt \
--attention_resolutions 32,16,8 --class_cond False --use_scale_shift_norm False --dropout 0.0 \
--image_size 256 --num_channels 256 --num_head_channels 64 --num_res_blocks 2 \
--num_samples 32 \
--resblock_updown True --use_fp16 True --weight_schedule uniform
My code looks like this:
model_kwargs = {}
sample = karras_sample(
diffusion,
model,
(args.batch_size, 3, args.image_size, args.image_size),
steps=args.steps,
model_kwargs=model_kwargs,
device=dist_util.dev(),
clip_denoised=args.clip_denoised,
sampler=args.sampler,
sigma_min=args.sigma_min,
sigma_max=args.sigma_max,
s_churn=args.s_churn,
s_tmin=args.s_tmin,
s_tmax=args.s_tmax,
s_noise=args.s_noise,
generator=generator,
ts=ts,
)
x_out, sample = iterative_inpainting(
distiller=model,
images=sample,
x=generator.randn(*sample.shape, device=dist_util.dev()), #* sigma_max
ts=ts,
t_min=0.002,
t_max=80.0,
rho=7.0,
steps=40,
generator=generator,
)
x_out = ((x_out + 1) * 127.5).clamp(0, 255).to(th.uint8)
x_out = x_out.permute(0, 2, 3, 1)
x_out = x_out.contiguous()
sample = ((sample + 1) * 127.5).clamp(0, 255).to(th.uint8)
sample = sample.permute(0, 2, 3, 1)
sample = sample.contiguous()
The masked images look ok.
@aarontan-git @arittenbach Could you find some improvement? The inpainting does not work properly, what am I doing wrong?
In Algorithm 4 in the paper, they described A as an invertible linear transformation, that maps images to the latent space.
I cannot see any mapping to the latent space.
from consistency_models.
Related Issues (20)
- The parameter sigma_data for KarrasDenoiser.
- Sampling with Diffusers HOT 1
- RuntimeError in the sample `diffusers` code in README.md HOT 2
- Different batches size while training
- “multi-gpu error” dist.all_gather(gathered_samples, sample) # gather not supported with NCCL HOT 1
- where can we get the ref_batches/imagenet64.npz file for FID evaluation HOT 1
- [inpainting] wrong shape?
- Request for unconditional model on ImageNet 64*64
- GPU Training Time for CD and CT Models on ImageNet64X64?
- Why does the training not stop?
- ModuleNotFoundError: No module named 'cm' HOT 3
- Why use teacher model in Consistency training HOT 5
- Can I use the checkpoint provided by the original EDM paper to train the consistency model?
- Is there a non-stochastic sampling method? HOT 1
- Issue with `use_fp16=True` Leading to Type Conversion Error in `unet.py` HOT 1
- An module error when running cm_train.py HOT 1
- cannot import name '_get_cpp_backtrace' from 'torch._C'请问这种问题该怎样解决呢 HOT 2
- How is Consistency for Any Point on the ODE Path Achieved in the Consistency Model?
- After running, no .npz file appeared HOT 2
- the libfabric EFA provider is operating in a condition that could result in memory corruption or other system errors. HOT 11
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from consistency_models.