What happened
Linum’s November 4 announcement describes a roughly 2-billion-parameter image generator trained on 660 million samples. The reported $21,000 is training compute, not the full cost of developing a commercial product. Code and weights are released under Apache 2.0.
The repository documents a CUDA workflow using Qwen3.5-4B as the text encoder. Listed weight sizes are 8.7 GB for the image model in fp32 and 8 GB for the encoder in bf16, before activations. Its five supported image shapes are around one megapixel. The default prompt-expansion step uses an external paid model, but users can bypass it with a detailed caption.
The release cannot reliably render text and is explicitly unsuitable for posters. Linum also identifies weaknesses in full-body scenes, groups and physical details. The cover reproduces the developer’s actual generated samples, not photographs or independent test results.
Why it matters
Releasing a reproducible checkpoint makes the efficiency claim testable. For creative teams, the practical trade-off is control over local inference against hardware, setup and quality-assurance work. An inexpensive training run alone does not establish low production cost or competitive video performance.
What to watch
Watch for independent reproductions of training cost and image quality, and a separately released video checkpoint.
Sources & verification
Source links are not a count of independent confirmations.
01Report a correction
Help us get it right
Include the disputed statement and, if possible, a supporting source URL. Please do not send personal or sensitive information.
