ali-vilab / ranni Goto Github PK

this method is powerful, i hope you can release more code and checkpoint

....
The conflict is caused by:
The user requested safetensors==0.2.7
transformers 4.30.2 depends on safetensors>=0.3.1

To fix this you could try to:

1. loosen the range of package versions you've specified
2. remove package versions to allow pip attempt to solve the dependency conflict

Hi, thanks for sharing this amazing work.
I'm currently experimenting with continuous editing. In the readme, it is specified that we should be able to modify the boxes by modifying the prompt. Unfortunately, this not looks immediate to me, can you provide more explanation?

Also, do you have an ETA for the release of the more powerful model? Thanks in advance!

Hi, I want to know, What is the minimum GPU memory required? I had inferenced the model in GPU RTX3090 24G, but it turned out of CUDA out of memory, what can I do to reduce the GPU memory ? Thanks.

Is there a way to use quantized models? The current version is really out of reach for most people with consumer-grade GPUs.

Also, will you train / release models for SD1.5 and SDXL?

Thanks!

I got good results using the recommended prompt and seed, but when I used the recommended prompt and random seed I got worse results.
Good Result:
input prompt: a black dog and a white cat
input seed: 15

Bad Result:
input prompt: 10 apples on the table
input seed: 15

Any solutions?

The paper mentioned you use llama-13B as the LLM model. But the provided version is 7B. Have you observed any performance drop because of the different LLM size. Will you plan to release the 13B version?

Hi, I tried to reproduce some of your examples and this is what I got:

While I understand that errors may occur, it looks to me that something is broken in my installation, am I correct? I have loaded LLaMA in bfloat16, could it be caused by that in your opinion?

In the config, you disable the clip by setting the version to None, So the initializing the clip encounter the following error. How to fix this bug?

cond_stage_config:
target: ranni.ranni.HackedFrozenOpenCLIPEmbedder
params:
freeze: True
layer: "penultimate"
version: null # disable loading CLIP here

size mismatch for base_model.model.model.layers.31.self_attn.q_proj.lora_B.default.weight: copying a param with shape torch.Size([4096, 64]) from checkpoint, the shape in current model is torch.Size([5120, 64]).

I met error like this, how do I solve it ?

Segmentation fault error when doing panel-to-image

Error also happens on sample images / generation

In the paper, you give the system prompt with detailed illustration of coordinate system and the example. But in this codebase, you remove those contents in the box prompt. Is there any difference in these settings?

after run the demo and select an example, this is what I got:

Thank you for work.
The error encountered during runtime is as follows.

It seems that the models is missed in ldm package