Practitioner distills Imagen 4 Disney-renaissance style to Illustrious 2.0 LoRA before June 30 sunset
A community member is training an Illustrious 2.0 LoRA to preserve Google's Imagen 4 Disney-renaissance style before the model sunsets, but early results show artifacts and inconsistent output quality.
Google's Imagen 4 and Imagen 4 Ultra are being sunset on June 30, and at least one practitioner is racing to preserve what they describe as the only models that reliably output a convincing 1990s Disney-renaissance look — the blurry-edge shading that defined the CAPS animation era.
The user trained a LoRA on Illustrious 2.0 using a 100-image dataset generated entirely by Imagen 4, all 16:9 at 1408×768. Training ran through TensorArt at 3 repeats and 10 epochs for 2,910 steps, with auto-labeling via wd-v1-4-vit-tagger-v2. The resulting LoRA captures the target style but introduces random artifacts, inconsistent line quality, and what the creator calls "the way AI gen looked like 2 years ago" — sloppy rather than polished. The training run cost nearly $10, and the user weighed whether to retrain with Imagen 4 Ultra images (sharper, 2× resolution, but $0.06 per image) or pre-process the existing dataset with denoising and upscaling.
After community feedback, the creator concluded that Illustrious 2.0 itself may be the bottleneck — "a very out of date model that's pretty far behind the curve." A secondary attempt with a Qwen Image Edit 2511 LoRA via FAL's trainer produced better character fidelity, though it unexpectedly edited backgrounds despite training only on character pairs. The user now plans to keep the Qwen LoRA and manually fix backgrounds rather than continue distilling to Illustrious.