AI & ML interests

None defined yet.

multimodalartΒ 
posted an update 3 months ago
view post
Post
12730
Self-Forcing - a real-time video distilled model from Wan 2.1 by @adobe is out, and they open sourced it 🐐

I've built a live real time demo on Spaces πŸ“ΉπŸ’¨

multimodalart/self-forcing
Β·
linoytsΒ 
posted an update 4 months ago
view post
Post
4673
FramePack is hands down one of the best OS releases in video generation πŸ™‡πŸ»β€β™€οΈπŸ€―
βœ… fully open sourced + amazing quality + reduced memory + improved speed
but more even - its gonna facilitate *soooo* many downstream applications
like this version adapted for landscape rotation πŸ‘‡https://huggingface.co/spaces/tori29umai/FramePack_rotate_landscape
  • 2 replies
Β·
linoytsΒ 
posted an update 4 months ago
multimodalartΒ 
posted an update about 1 year ago
radamesΒ 
posted an update over 1 year ago
view post
Post
7509
Thanks to @OzzyGT for pushing the new Anyline preprocessor to https://github.com/huggingface/controlnet_aux. Now you can use the TheMistoAI/MistoLine ControlNet with Diffusers completely.

Here's a demo for you: radames/MistoLine-ControlNet-demo
Super resolution version: radames/Enhance-This-HiDiffusion-SDXL

from controlnet_aux import AnylineDetector

anyline = AnylineDetector.from_pretrained(
    "TheMistoAI/MistoLine", filename="MTEED.pth", subfolder="Anyline"
).to("cuda")

source = Image.open("source.png")
result = anyline(source, detect_resolution=1280)
radamesΒ 
posted an update over 1 year ago
view post
Post
7038
At Google I/O 2024, we're collaborating with the Google Visual Blocks team (https://visualblocks.withgoogle.com) to release custom Hugging Face nodes. Visual Blocks for ML is a browser-based tool that allows users to create machine learning pipelines using a visual interface. We're launching nodes with Transformers.js, running models on the browser, as well as server-side nodes running Transformers pipeline tasks and LLMs using our hosted inference. With @Xenova @JasonMayes

You can learn more about it here https://huggingface.co/blog/radames/hugging-face-google-visual-blocks

Source-code for the custom nodes:
https://github.com/huggingface/visual-blocks-custom-components
radamesΒ 
posted an update over 1 year ago
view post
Post
2154
AI-town now runs on Hugging Face Spaces with our API for LLMs and embeddings, including the open-source Convex backend, all in one container. Easy to duplicate and config on your own

Demo: radames/ai-town
Instructions: https://github.com/radames/ai-town-huggingface
Β·
multimodalartΒ 
posted an update over 1 year ago
view post
Post
28502
The first open Stable Diffusion 3-like architecture model is JUST out πŸ’£ - but it is not SD3! πŸ€”

It is Tencent-Hunyuan/HunyuanDiT by Tencent, a 1.5B parameter DiT (diffusion transformer) text-to-image model πŸ–ΌοΈβœ¨, trained with multi-lingual CLIP + multi-lingual T5 text-encoders for english 🀝 chinese understanding

Try it out by yourself here ▢️ https://huggingface.co/spaces/multimodalart/HunyuanDiT
(a bit too slow as the model is chunky and the research code isn't super optimized for inference speed yet)

In the paper they claim to be SOTA open source based on human preference evaluation!
radamesΒ 
posted an update over 1 year ago
view post
Post
2648
HiDiffusion SDXL now supports Image-to-Image, so I've created an "Enhance This" version using the latest ControlNet Line Art model called MistoLine. It's faster than DemoFusion

Demo: radames/Enhance-This-HiDiffusion-SDXL

Older version based on DemoFusion radames/Enhance-This-DemoFusion-SDXL

New Controlnet SDXL Controls Every Line TheMistoAI/MistoLine

HiDiffusion is compatible with diffusers and support many SD models - https://github.com/megvii-research/HiDiffusion
  • 1 reply
Β·
pcuenqΒ 
posted an update over 1 year ago
view post
Post
8807
OpenELM in Core ML

Apple recently released a set of efficient LLMs in sizes varying between 270M and 3B parameters. Their quality, according to benchmarks, is similar to OLMo models of comparable size, but they required half the pre-training tokens because they use layer-wise scaling, where the number of attention heads increases in deeper layers.

I converted these models to Core ML, for use on Apple Silicon, using this script: https://gist.github.com/pcuenca/23cd08443460bc90854e2a6f0f575084. The converted models were uploaded to this community in the Hub for anyone that wants to integrate inside their apps: corenet-community/openelm-core-ml-6630c6b19268a5d878cfd194

The conversion was done with the following parameters:
- Precision: float32.
- Sequence length: fixed to 128.

With swift-transformers (https://github.com/huggingface/swift-transformers), I'm getting about 56 tok/s with the 270M on my M1 Max, and 6.5 with the largest 3B model. These speeds could be improved by converting to float16. However, there's some precision loss somewhere and generation doesn't work in float16 mode yet. I'm looking into this and will keep you posted! Or take a look at this issue if you'd like to help: https://github.com/huggingface/swift-transformers/issues/95

I'm also looking at optimizing inference using an experimental kv cache in swift-transformers. It's a bit tricky because the layers have varying number of attention heads, but I'm curious to see how much this feature can accelerate performance in this model family :)

Regarding the instruct fine-tuned models, I don't know the chat template that was used. The models use the Llama 2 tokenizer, but the Llama 2 chat template, or the default Alignment Handbook one that was used to train, are not recognized. Any ideas on this welcome!
Β·
radamesΒ 
posted an update over 1 year ago
view post
Post
2537
I've built a custom component that integrates Rerun web viewer with Gradio, making it easier to share your demos as Gradio apps.

Basic snippet
# pip install gradio_rerun gradio
import gradio as gr
from gradio_rerun import Rerun

gr.Interface(
    inputs=gr.File(file_count="multiple", type="filepath"),
    outputs=Rerun(height=900),
    fn=lambda file_path: file_path,
).launch()

More details here https://huggingface.co/spaces/radames/gradio_rerun
Source https://github.com/radames/gradio-rerun-viewer

Follow Rerun here rerun
radamesΒ 
posted an update over 1 year ago
radamesΒ 
posted an update over 1 year ago
radamesΒ 
posted an update over 1 year ago
radamesΒ 
posted an update over 1 year ago
view post
Post
2800
Following up on @vikhyatk 's Moondream2 update and @santiagomed 's implementation on Candle, I quickly put togheter the WASM module so that you could try running the ~1.5GB quantized model in the browser. Perhaps the next step is to rewrite it using https://github.com/huggingface/ratchet and run it even faster with WebGPU, @FL33TW00D-HF .

radames/Candle-Moondream-2

ps: I have a collection of all Candle WASM demos here radames/candle-wasm-examples-650898dee13ff96230ce3e1f
radamesΒ 
posted an update over 1 year ago
view post
Post
3854
Testing new pix2pix-Turbo in real-time, very interesting GAN architecture that leverages SD-Turbo model. Here I'm using edge2image LoRA single-step inference 🀯

It's very interesting how ControlNet Canny quality is comparable, but in a single step. Looking forward to when they release the code: https://github.com/GaParmar/img2img-turbo/issues/1

I've been keeping a list of fast diffusion model pipelines together with this real-time websocket app. Have a look if you want to test it locally, or check out the demo here on Spaces.

radames/real-time-pix2pix-turbo

Github app:
https://github.com/radames/Real-Time-Latent-Consistency-Model/

You can also check the authors img2img sketch model here

gparmar/img2img-turbo-sketch

Refs:
One-Step Image Translation with Text-to-Image Models (2403.12036)

cc @gparmar @junyanz
multimodalartΒ 
posted an update over 1 year ago
view post
Post
The Stable Diffusion 3 research paper broken down, including some overlooked details! πŸ“

Model
πŸ“ 2 base model variants mentioned: 2B and 8B sizes

πŸ“ New architecture in all abstraction levels:
- πŸ”½ UNet; ⬆️ Multimodal Diffusion Transformer, bye cross attention πŸ‘‹
- πŸ†• Rectified flows for the diffusion process
- 🧩 Still a Latent Diffusion Model

πŸ“„ 3 text-encoders: 2 CLIPs, one T5-XXL; plug-and-play: removing the larger one maintains competitiveness

πŸ—ƒοΈ Dataset was deduplicated with SSCD which helped with memorization (no more details about the dataset tho)

Variants
πŸ” A DPO fine-tuned model showed great improvement in prompt understanding and aesthetics
✏️ An Instruct Edit 2B model was trained, and learned how to do text-replacement

Results
βœ… State of the art in automated evals for composition and prompt understanding
βœ… Best win rate in human preference evaluation for prompt understanding, aesthetics and typography (missing some details on how many participants and the design of the experiment)

Paper: https://stabilityai-public-packages.s3.us-west-2.amazonaws.com/Stable+Diffusion+3+Paper.pdf
Β·
multimodalartΒ 
posted an update over 1 year ago