Welcome back. Let's get into it.
Black Forest Labs just dropped FLUX Virtual Try-On — and the headline number here is sub-4 seconds per generation. It's live on the BFL API and on fal right now. What makes this one interesting isn't just the speed — it's what it's actually good at. Logo preservation. Stitching details. Print patterns. The kind of fine texture that virtual try-on tools usually mangle completely. And it keeps person identity consistent whether you're doing a packshot or a full outfit shot. That's a real workflow tool for e-commerce teams, not just a demo.
And while we're on BFL — fal also added FLUX Erase to their platform. Single-image input, you draw a mask, and it handles object removal, text erasure, and cleanup. The key distinction here is that this operates at the model level — it's a genuine erase-and-reconstruct approach, not just an inpainting hack layered on top. Clean, focused, useful.
Switching over to Hermes Agent — version 0.15.0 just landed, and this release is packed. The big feature addition is direct Krea 2 API support through the Fal image gen provider. But the under-the-hood numbers are what caught my attention — 50% faster load times, and session search that is now 750 times faster than before. Seven-hundred-and-fifty times. This update came from 747 pull requests by 321 contributors. That's a genuinely large open-source push.
Okay — HiDream. This one has a bit of a story behind it.
A model called HiDream-O1-Image-Dev-2604 is currently sitting at the top of the Artificial Analysis Image Arena leaderboard for open weights models. What's interesting is that this model was previously released under a pseudonym — it was going by "Peanut." It's an 8 billion parameter base model with a prompt-enhancement pipeline baked in.
And speaking of Peanut — cocktailpeanut dropped a one-click local launcher for HiDream-O1-Image that runs on just 10 gigabytes of VRAM. The model uses a single unified transformer architecture and can take up to ten reference images at once for multi-subject editing. The example use cases being thrown around include things like — and I'm quoting here — "generate these guys having dinner" or "girl riding Shrek's back." Wild. But it works.
On the training side, ostris ran the first LoRA training on this model and flagged something worth paying attention to — it operates in pixel space directly. No VAE, no text encoder. That makes it architecturally pretty different from what people are used to, and the payoff is super-fast training. He was running it on a 5090 at batch size 2 without breaking a sweat.
And finally — Recraft. Two things worth flagging here.
First, someone shared a direct link to the fastest Recraft 4.1 demo that's publicly accessible — if you're evaluating image gen tools right now, that's worth a look.
Second — and this is the more technical one — there's a 6-step vectorization pipeline making the rounds built around Recraft. It goes: line drawing, into NB2 for 3D rendering, extract color codes, adjust angles in Photoshop, back into NB2 for flat rendering, and then finally into Recraft for the vector output. Six steps, but the results people are getting out of it are genuinely clean production-ready vectors. If you're doing illustration or design work, that pipeline is worth bookmarking.
That's your AI digest for 01 Jun 2026.