Ostris
AI / ML researcher and developer. https://ostris.com
ML at http://glif.app
- It kind of sucks that the AI/ML community seems to exclusively use twitter, at least people interested in the type of work I do.
- Flex.2-preview is here with text to image, universal control (line, pose, depth), and inpainting all baked into one model. Fine tunable with AI-Toolkit, Apache 2.0 license, 8B parameters. huggingface.co/ostris/Flex....
- HiDream LoRA fine tuning is now live on AI-Toolkit CLI and in the GUI. It currently requires a minimum of 36 GB of VRAM. Working on getting that down. github.com/ostris/ai-to...
- Flex Redux 512 was just released. SigLIP2 512 Vision Encoder. Works with Flex.1-alpha and FLUX.1-dev. Apache2.0 license. huggingface.co/ostris/Flex....
- AI generated nonsense music video with a LoRA I trained of myself (Wan2.1 14B). Prompts for video, video, and music is all AI generated. I edited it myself, that is the last step to automate for a fully automated AI slop machine. youtu.be/18SNWqdJt44
- Tutorial on how to train with targeted flow guidance with AI Toolkit youtu.be/OVhusDyWoZ4
- Made some long overdue ComfyUI nodes for Flex.1-alpha. A node to set guidance or bypass it for true CFG. LoRA loaders that automatically prune Flux LoRAs to work with Flex. They won't work perfect, but it should be decent for most use cases. github.com/ostris/Comfy...
- Wan 2.1 14B is amazing quality, but it is slow. The 1.3B version is extremely fast, and finetunes well. I trained a quick LoRA on it of myself for 1k steps. This is the most fun I have had messing with generative AI since the early SD1 days. Infinite personalized slop machine.
- First training sample montage of training a LoRA on Wan2.1 1.3B with AI Toolkit. Cruella. Still have to test my LoRA format to see if I can get it to load anywhere or if I need to modify it. Initial release will likely only support training on stills for now.
- Testing out the current training version of Flex.1-alpha/Flux.1-dev Redux adapter with SigLIP2 so400m 512. My Patreon supporters can download and use the current training version now. Public release coming soon when it is done cooking. youtu.be/J7zk9sURLcM patreon.com/posts/123794...
- Running a training test for training a redux adapter for Flex.1-alpha using siglip2-so400m-patch16-512. It is learning it remarkably fast. The 512 resolution should help with detail and texture vs the 384 v1 version.
- Flex.1-alpha face adapter training status update and current state demo. Still a long, long, looong way to go. youtu.be/7WmuH2_KuOc?...
- 5 days later and there was a UI. Only basic LoRA training for now. More features and tutorials coming soon.
- When a image/video gen model/method has a paper with no weights and no code. One can only assume that all the images/videos shown are heavily cherry picked.
- There is no joke funnier than OpenAI clutching their pearls because they suspect someone may have used their data to train a LLM without their permission. www.axios.com/2025/01/29/o...
- Reposted by OstrisIntroducing Kokoro.js, a new JavaScript library for running Kokoro TTS, an 82 million parameter text-to-speech model, 100% locally in the browser w/ WASM. Powered by 🤗 Transformers.js. WebGPU support coming soon! 👉 npm i kokoro-js 👈 Link to demo (+ sample code) in 🧵
- Reposted by Ostristhis is clearly just an amazon support human that is really into react
- Testing LoRA training for a new 8B model I have been cooking. Marty McFly and Pixar style LoRA training samples here. It is based on a pruned version of OpenFlux that has been continuously trained. I also trained a guidance embedding for it among other cool things.
- Has anyone had any luck converting FLUX LoRAs to SVDquant format? I have been trying to reverse engineer the process but keep hitting roadblocks.
- Reposted by OstrisA common question nowadays: Which is better, diffusion or flow matching? 🤔 Our answer: They’re two sides of the same coin. We wrote a blog post to show how diffusion models and Gaussian flow matching are equivalent. That’s great: It means you can use them interchangeably.
- Seriously, why is my Amazon Echo still dumber than a box of rocks? Have any of the home assistants evolved past the technology from 10 years ago? Is someone going to do something about this or do I need to?
- Finally! Google Calendar has dark mode!
- Testing training just an embedding that attaches like the Flux Redux output does. This is with 42 tokens doing cruella. It seems incapable of learning identity concatenating the embedding this way, leading me to think a face redux (which I am also training), may not be possible.
- You guys remember hyper networks? They just sort of disappeared when LoRA came along.
- Why don’t people want AI models to be trained on their ideals and views. AI models WILL develop a bias based on their training data. I would prefer the biases of the people on this platform to be trained into them. Otherwise, future AI will just be a crypto bro. Please support scraping public data.
- Reposted by OstrisHere, have PSGD-Kron and SOAP with FSDP2 support. Please go wild with it, let's see something finally replace ADAM. github.com/ethansmith20...
- OMW to Disney world for a few days with the fam. If you need something from me, wait.

- Being able to use my own domain as my handle is pretty cool.
- This image popped up in a training sample for a fine tune of redux I am cooking. I don't know what it is, but I want one,
- It is supposed to be someone wearing this shirt.
- I think I like this version of "The Scream" better. Fine tuning. Sample from fine tuning Flux Redux adapter.
- OMG. I love Flux redux with multiple images. Try it out here -> glif.app/@HighDruidMo...
- The Flux.1 dev redux adapter is pretty amazing. It is just 2 linear layers after siglip, but it is capable of accurately reproducing the content of the image and the text, even the misspelled parts. That is incredible. -> huggingface.co/black-forest...
- @bsky.app Please add bookmarks. I desperately need bookmarks.
- I think I have full fine tuning of FLUX on a 4090 at <3s/iter worked out. Still running a few tests, but it will be live in ai-toolkit soon.
- It is crazy how well Doing img2img with IC LoRA works for injecting a logo into an image. The game has changed. Try it out here -> glif.app/glifs/cm3o7d...
- Reposted by Ostris[This post could not be retrieved]