Handle images with MiniMax H3.

extensions_built_in/diffusion_models/minimax_h3/minimax_h3.py

@@ -1,7 +1,9 @@
 """MiniMax-H3 (33B joint video+audio DiT) for ai-toolkit.
 
 Supports t2v (t2va) and first-frame i2v (fl2va) training and sampling, with
-joint audio when the dataset provides it. The architecture lives in ./src/:
+joint audio when the dataset provides it. Image datasets train as single
+latent frames (keyframe-row geometry) and sampling with num_frames 1 renders
+a single image. The architecture lives in ./src/:
 
   - transformer.py: packed-sequence DiT, weight-compatible with the original
     ``MiniMaxAI/MiniMax-H3`` checkpoint keys

extensions_built_in/diffusion_models/minimax_h3/src/pipeline.py

@@ -86,12 +86,13 @@ class MiniMaxH3Pipeline:
         is_video = num_frames > 1
         if is_video:
             num_frames = packing.align_num_frames_down(num_frames)
+            t_lat = packing.video_latent_num_frames(num_frames)
         else:
-            # a single image still runs through the video path with the
+            # true single-frame generation (image mode, LTX-2.3 style): one
-            # minimum 5-frame clip; the first frame is returned
+            # latent frame with keyframe-row geometry — the same layout image
-            num_frames = packing.LATENTS_PER_CHUNK
+            # datasets train with, so image LoRAs sample in-distribution
-
+            num_frames = 1
-        t_lat = packing.video_latent_num_frames(num_frames)
+            t_lat = 1
         h_lat = height // 16
         w_lat = width // 16
         a_lat = packing.audio_latent_num_frames(num_frames)

extensions_built_in/diffusion_models/minimax_h3/src/vae.py

@@ -777,7 +777,13 @@ class MiniMaxH3VideoVAE(nn.Module):
         )
         with ctx:
             if z.shape[2] == 1:
-                dec = self._decode_clip(z)[:, :, -1:]
+                # a lone temporal token is OOD for the chunk-trained ViT
+                # decoder (visibly patchy output, ~16 dB vs ~35 dB roundtrip).
+                # Decode it as the first latent of a 2-latent clip instead:
+                # the first latent of a chunk covers exactly pixel frame 0 in
+                # the (1, 4, 4, 4, 4) grouping, so that frame is a clean,
+                # in-distribution single-frame decode.
+                dec = self._decode_video(torch.cat([z, z], dim=2))[:, :, :1]
             else:
                 dec = self._decode_video(z)
 

ui/src/app/jobs/new/options.tsx

@@ -764,7 +764,8 @@ export const modelArchs: ModelArch[] = [
         <p>
           Supports t2v and first-frame i2v (ctrl img / i2v datasets) with joint audio. The model is guidance-distilled —
           keep guidance scale at 1. Video is fixed 24 fps and frame counts snap down to the 17n+5 grid (5, 22, 39, 56,
-          ..., 107, 124 ≈ 5s).
+          ..., 107, 124 ≈ 5s). Image datasets (num_frames 1) train as single frames, and a sample with num_frames 1
+          renders a single image.
         </p>
       </div>
     ),