For text-to-video generation, disabling enable_PE can improve video quality.
#15
by fengjunwei - opened
Many workflows have enable_PE enabled by default, which invokes a lightweight version of the text encoder. This can lead to a loss of image details and other quality issues.
Although the quality still doesn't quite match the official API, using qwen3vl_32b_h3_ultra_uncensored_heretic_bf16.safetensors could help further narrow the gap.
Use 16MP. In my tests, generating a single image with a file size of 20MB takes only a few seconds. You can zoom in to see even more details.
I don't think any open-source model currently available can achieve this level of detail. Even closed-source models, when run locally, fail to match the level of detail this model delivers.