PrunaAI p-h3-image LoRA on our H3

The 2 GB p-h3-image LoRA (huggingface.co/PrunaAI/P-H3-Image-Comfy) on the bot's fl2va fp8 DiT, with Pruna's settings: bypass LoRA loader at 1.0, euler on its 8 sigmas, shifts 1/1, CFG 1, no prompt upsampler. Four ways to feed it: our H3 text encoder or Pruna's Qwen3-VL-4B one, and the SOL VAE or Pruna's single-frame VAE (Pruna's parts load through the ph3_image custom node, a port of ComfyUI commit 576b642e). Route 1 runs our encoder through MiniMaxH3ReferenceToVideo at length 1 like the bot (no extra files or node); the other cards use Pruna's p-h3-image Conditioning node (Pruna's VAE only decodes the single latent frame that node makes). 2 MP (1440×1440 or the source's aspect), seed 42. The blue card is the bot's H3 Pika result for the same prompt and seed (older run), the teal card Pika + chroma 1.0 (bot model h3chroma), rendered now with the same prompts as the Pruna cards (no Task: opener). Click a picture to open the viewer: ← / → flips between versions, Z or click toggles 1:1 pixels.

H3 · Pika (the bot today) H3 · Pika + chroma 1.0 (bot model h3chroma) Pruna LoRA · our encoder + SOL VAE (route 1) Pruna LoRA · our encoder via Pruna's node + SOL VAE Pruna LoRA · our encoder via Pruna's node + Pruna VAE Pruna LoRA · Pruna encoder + Pruna VAE (route 2, as shipped) Pruna LoRA · Pruna encoder + SOL VAE source picture