Pinned Tweet
If you appreciate my contributions to open source software and models, please consider becoming a sponsor. Supporting open source projects and their creators is vital for open innovation, and financial support from people like you makes my work possible.
12
20
188
69,099
Ostris retweeted
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
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
9
3
26
3,434
I think there is a better training target for flow-matching diffusion models. I call it x0 velocity.
It keeps the forward process and the sampling trajectory of flow matching exactly. The network's output is converted back to a velocity with one subtraction, so any flow-matching guidance method runs unchanged. What changes is the target the network regresses onto:
target = noisy - (noise - clean) = clean + (1 - t) * (clean - noise)
At the pure-noise end this is plain x0 prediction. Below that it is x0 extrapolated one unit along the flow line. The output is image-shaped at every timestep, never a noise field, and the correction the network has to make, the difference between its input and its target, has the same magnitude at every timestep. Plain x0 prediction's correction shrinks to zero at the clean end, which starves the fine-detail steps. Velocity prediction's correction grows toward the clean end, where the network has to reproduce and invert its own input. x0 velocity sits between them: a constant-size correction with an image-like output, and a loss that is exactly the flow-matching velocity loss, so it needs no timestep weighting.
In my testing it trains more stably than either, and it is more tolerant of high-frequency content, which lets models perform better with more latent channels and deeper compression. 🧵
It keeps the forward process and the sampling trajectory of flow matching exactly. The network's output is converted back to a velocity with one subtraction, so any flow-matching guidance method runs unchanged. What changes is the target the network regresses onto:
target = noisy - (noise - clean) = clean + (1 - t) * (clean - noise)
At the pure-noise end this is plain x0 prediction. Below that it is x0 extrapolated one unit along the flow line. The output is image-shaped at every timestep, never a noise field, and the correction the network has to make, the difference between its input and its target, has the same magnitude at every timestep. Plain x0 prediction's correction shrinks to zero at the clean end, which starves the fine-detail steps. Velocity prediction's correction grows toward the clean end, where the network has to reproduce and invert its own input. x0 velocity sits between them: a constant-size correction with an image-like output, and a loss that is exactly the flow-matching velocity loss, so it needs no timestep weighting.
In my testing it trains more stably than either, and it is more tolerant of high-frequency content, which lets models perform better with more latent channels and deeper compression. 🧵
11
8
138
6,185
I am converting MOSS image into a 64x latent compression model. I am applying a patch 8 to the latent instead of the patch 2 it has. It is still training, these are both 1 day on a single GPU, but here are current samples. First is flow-matching, second it x0 velocity.
0
0
15
1,182
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.
6
1
78
4,508
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.
10
2
107
4,498
Ostris retweeted
moss base 8 steps DMD tune huggingface.co/lodestones/MOS…
moss SFT 8 steps DMD tune huggingface.co/lodestones/MOS…
go wild!
moss SFT 8 steps DMD tune huggingface.co/lodestones/MOS…
go wild!
🤗 MOSS-Image is now open source on @huggingface .
Built on a 5B Diffusion Transformer (DiT), MOSS-Image brings Chinese & English text rendering and controllable layouts to image generation — from posters and infographics to creative visuals.@Open_MOSS
What it offers:
📝 Long promup to 4K-token prompts, including multi-line text
📐 Experimental JSON layout control — specify text, objects, and their positions using bounding boxes
🖼️ Flexible generation — multiple aspect ratios, resolutions, and customizable inference settings 🛠️ Developer workflows — Python (Diffusers), Gradio, ComfyUI, and a Prompt Enhancer Skill
📦 Base & SFT checkpoints are both available, with open model weights and Apache-2.0 licensed inference code.
🙌 Huge thanks to @sgl_project @vllm_project @ModelScope2022 for the Day-0 support!
4
6
99
6,174
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.
36
18
584
23,516
Ostris AI Toolkit has day 1 support for the new NVIDIA RTX Spark. This laptop is absolutely amazing!
NVIDIA RTX Spark is here.
Build AI locally, bring your creative ideas to life, and take your favorite games along.
A new beginning for @Windows PCs starts today.
4
1
69
4,785
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.
9
11
262
17,837
MOSS Image Base support added as well. huggingface.co/OpenMOSS-Team/…
0
0
12
695
Ostris retweeted
goated team!
apache 2 and that's a lot of data !
3.5B trained into one model
huggingface.co/OpenMOSS-Team/…
apache 2 and that's a lot of data !
3.5B trained into one model
huggingface.co/OpenMOSS-Team/…
1
18
125
8,683
Ostris retweeted
Higgsfield promised to open-source their entire platform. They didn’t.
Let's help them.
For the next two weeks, Comfy and our partners are putting up $25,000 in prizes for the best open-source builds on the Comfy Developer Platform.
Partners: @NVIDIAAI, @runpod, @MiniMax_AI, 0K grand prize
→ RTX 5090: Most Innovative
→ RTX 5090: Most Viral
→ RTX 5090: Open Weights Only
→ 10 runners-up: $250 in Comfy credits
Let's help them.
For the next two weeks, Comfy and our partners are putting up $25,000 in prizes for the best open-source builds on the Comfy Developer Platform.
Partners: @NVIDIAAI, @runpod, @MiniMax_AI, 0K grand prize
→ RTX 5090: Most Innovative
→ RTX 5090: Most Viral
→ RTX 5090: Open Weights Only
→ 10 runners-up: $250 in Comfy credits
205
372
4,632
639,654
Ostris retweeted
1/8 Does text-to-image generation need language pretraining?
Today, we share Halcyon, a family of early-fusion, base models trained from scratch. Halcyon outperforms strong late-fusion baselines with up to 275× less pretraining compute. Models and recipes are open.
Halcyon is ram at Radical Numerics on the science of multimodal pretraining and scaling.
Today, we share Halcyon, a family of early-fusion, base models trained from scratch. Halcyon outperforms strong late-fusion baselines with up to 275× less pretraining compute. Models and recipes are open.
Halcyon is ram at Radical Numerics on the science of multimodal pretraining and scaling.
4
31
178
75,652
Ostris retweeted
It's time. It's fucking time!
Flux.3 Image Klein is out and we bring you full support for it in Diffusers.
Catch all the details in our post:
huggingface.co/blog/flux-3
Flux.3 Image Klein is out and we bring you full support for it in Diffusers.
Catch all the details in our post:
huggingface.co/blog/flux-3
8
12
89
10,311