Sol and Claude Code build 3D scenes from reference images

• Reference-to-3D is already a live Sol workflow, with gokayfem's demo generating a Three.js object from four views and gokayfem's prompt asking Claude Code to critique geometry, textures, lighting, and connection points.
• Video is becoming a design input, not just output: MengTo's walkthrough framed one-shot video-to-HTML as Sol's superpower, and MengTo's reply called it the new screenshot-to-HTML for animations and interactions.
• Blender workflows are moving toward natural-language scene control, with MengTo's Blender MCP demo saying zero Blender experience was enough when paired with skills and the Blender MCP repo.
• The build burst already spans Blender hardware, LEGO animation, Three.js characters, voxel cities, and screenplay revision loops from minchoi's MacBook example, minchoi's Three.js character, bilawalsidhu's voxel city, and rainisto's screenplay loop.

OpenAI's launch post says GPT-5.6 is generally available, with Sol as the flagship model and ultra as the highest-capability setting for parallel workstreams. gokayfem's exact prompt made the agent care about contact points: no floating parts, matching materials, matching lighting in the prompt. MengTo's Skills repo turns screen recordings into super prompts for one-shot HTML. BlenderMCP lets Claude create and inspect Blender scenes over MCP, while HyperFrames renders HTML, CSS, media, and seekable animations into deterministic MP4.

Gokayfem started from a multi-view reference image and asked GPT 5.6 Sol to produce one Three.js object, then wrap the result in a HyperFrames explanation video. The first note called it "10 minutes with 5.6 Sol Ultra" in gokayfem's post.

The stack in gokayfem's follow-up was compact:

• Reve 2.1 for reference images
• Three.js for the 3D representation
• HyperFrames for the video
• Claude Code on CLI as a critic, according to gokayfem's full prompt

The workflow is the interesting part: reference image as spec, code agent as modeler, Claude Code as critic, HTML video renderer as packaging.

Video went from inspiration to input in MengTo's workflow. MengTo's walkthrough called one-shot video-to-HTML Sol's real superpower and linked the open-sourced skills behind it.

The MengTo/Skills repo describes Video to Super Prompt as a skill that turns a screen recording of a design, landing page, or animation into a detailed prompt for one-shot HTML.

His replies tightened the claim:

• Video-to-HTML is the new screenshot-to-HTML for animations and interactions, according to MengTo's reply.
• Sol is probably similar to another model but better at design, in MengTo's comparison reply.
• The repo hit 2,000 stars in a week, according to MengTo's stars reply.
• Output variance got the most honest answer in MengTo's “stars aligned” reply.

Blender MCP is the other half: natural-language control of a live 3D tool. MengTo's Blender MCP demo said zero Blender experience plus skills was enough to create intricate, animated models.

The BlenderMCP repo says it connects Claude AI to Blender through MCP for prompt-assisted 3D modeling, scene creation, scene inspection, and object manipulation. It also lists viewport screenshots, Sketchfab downloads, Poly Haven assets, Hyper3D Rodin generation, remote hosts, and anonymous tool telemetry.

MengTo's own reaction was short: the detail level had "upgraded tremendously" in his reply.

Minchoi gathered ten examples less than five days after launch in his GPT-5.6 roundup. The creator-relevant ones clustered around 3D, animation, and UI cloning:

• A floating MacBook in Blender, from minchoi's MacBook example.
• 7,500 LEGO bricks assembling the Millennium Falcon, from minchoi's LEGO example.
• A rigged, animated 3D character running in Three.js, from minchoi's character example.
• A Blender scene built with almost no human help, from minchoi's Blender example.
• A screen recording turned into a polished app UI, from minchoi's app UI example.
• Developers noticing Sol inside Claude Code, from minchoi's Claude Code example.

Outside that roundup, bilawalsidhu's voxel city prompt asked Sol Ultra to recreate San Francisco as a 3D voxel city, and MengTo posted another Sol Ultra-made build with "What a time to be a builder."

Rainisto used Sol for writing, not visuals: one agent revised a screenplay, another read it cold with the same evaluation criteria, and the human picked directions between rounds.

Scores moved from 79 to 90 over a few hours, with a caveat: repeated independent agents varied by about 2 to 3 points, while the written notes stayed aligned, according to rainisto's follow-up. Earlier in the same thread, rainisto used MCP Blender to block a primitive Miami night street, then asked gpt-image-2 to turn it into a polished Miami nightlife image.