SolarRenderer: Few-Step Generative Reference Rendering
Python
docs: render.py describes both LoRAs merged into the weights
Enderfga0c381ea6ec12
diff --git a/scripts/render.py b/scripts/render.pyindex a48e2a1..a231e50 100644--- a/scripts/render.py+++ b/scripts/render.py@@ -14,8 +14,8 @@The inputs are encoded exactly as the training samples were (tools/omni_preencode.py): the structure video laid overevery target frame, the first frame over latent 0, the prompt through the released Qwen3-VL presentation. The-renderer is MiniMax-H3's Ref2VA transformer with the four-step LoRA merged into its weights and the control LoRA on-top, sampled in four evaluations with fresh noise between them. Output: 1344x768 (or 768x1344), 24 fps, 158 frames.+renderer is MiniMax-H3's Ref2VA transformer with the four-step LoRA and the control LoRA merged into its weights at+once, sampled in four evaluations with fresh noise between them. Output: 1344x768 (or 768x1344), 24 fps, 158 frames.With --causal the causal pair renders block-causally: each block of at most 5 latents sees the prompt, the first frame,the structure of the last 6 blocks and the clean frames of the 5 blocks before it - never a later block - and issampled in four evaluations before it is committed to the cache.@@ -132,8 +132,8 @@ CAUSAL = {"mode": "teacher", "chunk": 5, "window": 6, "corrupt": 0.0} # the caudef load_renderer(h3: str, control: str, fewstep: str, device: str, causal: bool = False, frames: int = FRAMES):- """MiniMax-H3's Ref2VA transformer, the four-step LoRA merged in, the control LoRA on top; returns a renderer for- clips of up to ``frames``."""+ """MiniMax-H3's Ref2VA transformer with the four-step and control LoRAs merged in; returns a renderer for clips of+ up to ``frames``."""import torch