Open sourcing InScene + InScene Annotate - a pair of LoRAs for steering in-scene image generation w/ QwenEdit. You can use a combination of prompts + annotations to generate consistent images inside a scene. In beta but v. powerful. Model + workflow below, training data soon.

The next 6 months of open source AI art are going to be so magical So many foundational elements - edit, music, 3d - are being unlocked So many powerful pieces to combine What a time to be alive

What people can do by combining various pieces of the Wan ecosystem is so impressive This uses two models that were released 6+ months apart - SCAIL for pose tracking w/ Uni3c for camera motion:

I tried GPT-Image again and it got something somewhat coherent but around 50% of the time I get the above garbled nonsense:

Same prompt, NBP vs. GPT Image 1.5:

Being mean to LLMs is the new being mean to waiters It hurts another being, erodes your soul, and eventually someone else will see in and label your forever Better to me mean to opponents on chess dot com

Trellis2 and SAM Audio today, new HunyuanWorld today Shaping up to be a good week for open models! If z-image edit and qwen music drop before Christmas I’ll be a very happy boy

The Sun is an enormous, free light source in the sky. It is super dumb to make tiny light sources on Earth. Stop wasting money on candles, lightbulbs, and flashlights.

I remember wondering if we would ever have this level of control back in the warpfusion days. One-to-All Animation: Alignment-Free Character Animation and Image Pose Transfer - https://t.co/J9sdqY80UH workflow in comments

"Under the hush of a winter night, this PR bestows Qwen Image Edit with enigmatic new powers"

One other thing GPT5.2 feels best as it vision understanding It’s the only model that gives me feedback and ideas from visual inputs that actually makes sense But it is ironically terrible at actually acting on what it sees, Gemini still feels best for tricky front end code

This is actually a very positive review, it's saved me a tonne of time as a second opinion on bugs + implementations. Probably 20 mins in just a few hours of usage. Recommend so far!

Also spotted about of details in an Opus-authored security review that prevented a bunch of back and forth

Leaning towards it as a second opinion on complex plans, feels overwhelmingly worse than Opus for most practical usage

2 anecdotes from GPT 5.2: 1) I got it to review a plan that Opus struggled with. It spotted the issue w/ it - would've saved me 30 mins 2) I asked for help on a bug & it tried a fix. I explain why that's probably not it. It agrees, thinks for 5 mins, & then tries the same fix

Google search volume for Sora vs. ChatGPT:

How my day is going:

(someone told me that they're often confused about whether i'm joking or not so i should clarify that this is half a joke)

This is actually good modern software engineering If you give an LLM more precise language - e.g. 'Condense the design of this section' - it takes it too literally and goes too far that direction Casually-phrased instruction leads to less autistic interpretation

PSA: there’s a truly excellent collection of Phantom Thread gifs available on your gif provider of choice: