Ostris
@ostrisai
AI / ML researcher and developer.
Creator of AI Toolkit -
Support my work -
350 Following 15.4K Followers
First experiment with Primal VAE: Training MOSS Image to use the 2x1 with a spatial compression of 64x. These are the current results from an overnight run. With this compression, it generates natively at 4096x4096 for same compute of 1024x1024 on MOSS Image.
Show more
Today I am releasing Primal VAE in 2 variants, 2x1 and 2x2 (compression)x(channels). It only has a 2x compression and 1-2 channels, but with patching, it matches the compression of the FLUX 1 & 2 VAEs and beats FLUX2 VAE on reconstruction. More in README.
Show more
Today I am releasing Primal VAE in 2 variants, 2x1 and 2x2 (compression)x(channels). It only has a 2x compression and 1-2 channels, but with patching, it matches the compression of the FLUX 1 & 2 VAEs and beats FLUX2 VAE on reconstruction. More in README.
Show more
The thing about PRs in the age of coding agents is that I can build almost any feature or fix any bug with significantly less effort than it takes to read over and test a PR. There are exceptions, but I think the way we manage open source projects is going to have to change.
Show more
kale-img-1 fine-tuning results
1. @ostrisai / 8.10 / moss
2. @AmandineFlachs / 7.61 / moss
3. @suhaaspk / 7.20 / moss
4. @Assu_Bhadhu / 6.90 / moss
5. @max_pe2002 / 5.35 / hdo1
6. @hasnain__dev / 3.51 / sdxl
(same prompt, seed, resolution)
Show more
paying $5k to whoever fine-tunes kale-img-1, an img model for kids
dm me
i'll send u the dataset
u pick a base model we can license commercially
u comment with ur result here
ur result gets judged on:
- recreating 5 imgs from the dataset almost 1:1
- making 5 new imgs in our style
Show more
Ostris AI Toolkit now supports @bria_ai_ Fibo 2.0
You guys have no idea how many websites I have and maintain that use PHP and Wordpress that are affected by this. I am so sick of running and testing updates.....
Effective Immediately, all organizations using PHP for their systems are advised to be prepared for an update addressing two critical vulnerabilities. PHP is widely used for web applications. These vulnerabilities often provide a direct path to the server's file system or database. If an immediate update is not feasible, it is recommend that organizations take the system offline until it can be fully mitigated. NSA will follow up shortly with more specific guidance.
Show more
I made this 6-7 years ago using some style transfer methods I was working on at the time. I spent months of tinkering to get this level of temporal consistency. We have come so far so fast. I cannot even imagine where things will be in another 6 years at this pace.
Show more
Added LoRA training for Qwen 3.8 27B in AI Toolkit and spent the morning optimizing the captioner for it. It now captions at 70-80 tok/s with 1024 res image inputs and a decently long system prompt on a downclocked 5090 using less than 20GB of vram.
Show more
moss base 8 steps DMD tune
moss SFT 8 steps DMD tune
go wild!
I think one of the reasons people love MOSS Image so much is they intentionally removed AI images from their training data. There is a clear trend of new models training on AI images and having an overly ChatGPT look to them. Moss just looks better because they didn't do that.
Show more
So @Open_MOSS MOSS Image learns image editing fast. It picked up my "Make this person a cyclops" concept almost immediatly. I added support to train it as an edit model in AI Toolkit. I am training a style reference LoRA now.
Show more
Ostris AI Toolkit has day 0 support for FLUX 3 Image Klein. Thank you @bfl_ai for the early access and the amazing model!
Ostris AI Toolkit now supports @Alibaba_Qwen Qwen-Image-2.1
Here are some of the current samples with the default prompts in AI Toolkit generated using the current state of the model. The AR model has clearly grasped how to create images that match the prompts.
Show more
you don't need to pay too much attention!
full attn is overkill and should be used sparingly
calibrated krea2 from full attn on 48 heads into 12 heads full attn while the rest is scanline attn (band diagonal attn).
scanline attn is block sparse by design (see the gif)
Show more
Captioning music is difficult. The LLM has to be greedy because you need to be precise, but this makes it susceptible to fall into loops. Lyrics often repeat over and over, which makes it want to loop more. You can penalize this, but that breaks real loops.
Show more
Added initial support to fine tune LLMs in AI Toolkit with a focus on Omni models. First one added is
Qwen2.5-Omni-7B and I trained a LoRA for it to better caption and transcribe music files. You can use these LoRAs in the dataset captioner along with default cloud ones.
Show more
I made an explanation video for anyone curious to know more about the specifics without diving into gritty details.
This is an autoregressive model and fine tuning it is still highly experimental. Let me know any issues you run into. PR is here
