Seedream 5.5 generates transparent assets, survives 100 edit rounds
The new model nails dense small text (0.4% corruption) too — while Lightricks open-sources a 4K video+audio model and Liquid ships a 1GB on-device agent.

Copy markdown
Drop-in transparent assets, no background removal
ByteDance Seed's new image model outputs native RGBA with a real alpha channel, so generated elements — cutout product shots, icons, logo marks, characters — paste straight into a layout with no separate background removal. It also preserves unedited regions at the pixel level across a demoed 100 edit rounds and cuts severe text corruption to 0.4%. Live now in the Dola app and BytePlus Playground, with a Pro tier to follow.
Lightricks open-sources a 4K video+audio model
LTX-2.5 generates synchronized video and audio from text, image, or video, with native multishot so characters and style persist across cuts — up to 3840×2176 through its upscaling pipeline. The weights are free for commercial use under $10M revenue and run locally with fp8 casting and CPU offload, in ComfyUI/Diffusers, and most LTX-2.3 LoRAs reportedly still work.
A 1GB agent model that runs on your phone
Liquid AI's LFM3-1.7B is a 1.73B tri-hybrid model with a 256K context that fits in about 1GB at 4-bit — benchmarked on an iPhone 17 Pro, an M5 Max, and a Ryzen AI Max+ 395. It's post-trained for tool calling and structured output, ships GGUF/MLX/ONNX for Ollama and llama.cpp, and fine-tunes on a single GPU; skip it for heavy coding or multi-step reasoning.