Add lightx2v 8-step v1.0 LoRA as lightx8; fix step-count off-by-one

lightx8 loads minimax_h3_fl2v_turbo_8step_v1.0_bf16.safetensors through
the same diffusers-native PEFT mapping (rank 128, alpha 8, scale 0.0625),
suggested 8 steps, H3_LIGHTX8=off skips it.

MiniMaxH3Scheduler reads num_inference_steps as sigma grid points
including the terminal zero, so N evaluations need N + 1 points (as the
ModelTC inference script does). Pass steps + 1 so the UI's step count
matches the actual NFE count.

README.md

@@ -55,6 +55,10 @@ and the low-rank factors stay resident so switching is an in-place unfold/fold t
 `H3_LORA` selects the larry file (`off` skips it), `H3_LIGHTX=off` skips lightx, `H3_LORA_DEFAULT` picks which set
 starts folded, and `H3_LORA_STRENGTH` scales the larry update (the card's sharpness/artifact dial).
 
+The lightx set also comes as `lightx8` — the 8-step v1.0 diffusers checkpoint
+(`minimax_h3_fl2v_turbo_8step_v1.0_bf16.safetensors`), same PEFT layout and scale, suggested 8 steps.
+`H3_LIGHTX8=off` skips it.
+
 A third, non-turbo set is [`fal/MiniMax-H3-Realism-People-LoRA`](https://huggingface.co/fal/MiniMax-H3-Realism-People-LoRA)
 (`realism` in the dropdown, trigger word `r34l1sm`): realistic people at the full 28 steps. It ships in the same
 reference-tree layout as larry under a `diffusion_model.` prefix, so it folds through the same key mapping.
@@ -191,7 +195,8 @@ one-time `PIPE.to("cuda")` is inside the first row's 339 s and does not reappear
 | `H3_LORA` | `minimax_h3_turbo_v4_step600_ema.safetensors` | Turbo LoRA file folded into the transformer at startup. `off` disables. |
 | `H3_LORA_REPO` | `larryvrh/MiniMax-H3-Turbo-Lora` | Hub repo the LoRA is fetched from. |
 | `H3_LORA_STRENGTH` | `1.0` | Scales the larry LoRA delta (sharpness/artifact trade-off). |
-| `H3_LIGHTX` | `on` | Set to `off` to skip loading the lightx2v LoRA set. |
+| `H3_LIGHTX` | `on` | Set to `off` to skip loading the lightx2v 4-step LoRA set. |
+| `H3_LIGHTX8` | `on` | Set to `off` to skip loading the lightx2v 8-step v1.0 LoRA set. |
 | `H3_REALISM` | `on` | Set to `off` to skip loading the fal realism-people LoRA set. |
 | `H3_JOYFOX` | `on` | Set to `off` to skip loading the joyfox 4-step turbo LoRA set. |
 | `H3_LORA_DEFAULT` | `larry` | Which loaded LoRA set starts folded (`larry` / `lightx` / `realism` / `joyfox`). |

app.py

@@ -254,6 +254,8 @@ def _generate(prompt_embeds, text_token_tags, image, last_image, height, width,
 
     # The modular denoise loop has no callback hook, so count video scheduler steps through a wrapper: one call per
     # denoise step. When the count reaches the total, what remains inside `PIPE(...)` is the two decoders.
+    # `MiniMaxH3Scheduler` reads `num_inference_steps` as sigma grid points, terminal zero included, so N model
+    # evaluations need N + 1 points (the lightx/ModelTC inference script does the same).
     steps = int(steps)
     PROGRESS.update(phase="denoise", step=0, steps=steps)
     original_step = PIPE.scheduler.step
@@ -282,7 +284,7 @@ def _generate(prompt_embeds, text_token_tags, image, last_image, height, width,
             height=height,
             width=width,
             num_frames=num_frames,
-            num_inference_steps=steps,
+            num_inference_steps=steps + 1,
             generator=torch.Generator("cpu").manual_seed(int(seed)),
         )
     finally:
@@ -326,8 +328,8 @@ def _fit_keyframe(image_path, current_canvas):
 
 
 def _resolve_lora(lora, use_lora) -> str:
-    """`lora` (`larry` / `lightx` / `realism` / `joyfox` / `off`) wins; the legacy bool maps onto `larry` / `off`."""
-    if isinstance(lora, str) and lora in ("larry", "lightx", "realism", "joyfox", "off"):
+    """`lora` (`larry` / `lightx` / `lightx8` / `realism` / `joyfox` / `off`) wins; the legacy bool maps onto `larry` / `off`."""
+    if isinstance(lora, str) and lora in ("larry", "lightx", "lightx8", "realism", "joyfox", "off"):
         return lora
     return "larry" if use_lora else "off"
 
@@ -452,7 +454,8 @@ def studio_config():
         # The LoRA dropdown: value -> {label, suggested steps}.
         "loras": {
             **{
-                name: {"label": spec["label"], "steps": {"larry": 6, "lightx": 4, "realism": 28, "joyfox": 4}.get(name, 6)}
+                name: {"label": spec["label"],
+                       "steps": {"larry": 6, "lightx": 4, "lightx8": 8, "realism": 28, "joyfox": 4}.get(name, 6)}
                 for name, spec in sets.items()
             },
             "off": {"label": "off (base model)", "steps": 28},

h3_lora.py

@@ -17,9 +17,10 @@ The two supported LoRAs ship in different layouts:
   `final_layer.adaln_proj.linear` -> `norm_out.linear`. The row transforms are applied to `lora_B` directly
   (rows of `B @ A` are rows of `B`), so no full delta is ever materialized at load.
 
-* `lightx` (`lightx2v/Minimax-h3-Turbo`) is a PEFT checkpoint against the diffusers tree itself —
+* `lightx` / `lightx8` (`lightx2v/Minimax-h3-Turbo`) are PEFT checkpoints against the diffusers tree itself —
   `transformer_blocks.N.attn.to_q.lora_A.default.weight` and friends — rank 128, `alpha == 8`, so the fold scale is
   `8 / 128 = 0.0625` (matching `set_adapters(weights=1.0)` in their inference script). Keys map name-for-name.
+  `lightx` is the 4-step file, `lightx8` the 8-step v1.0 file.
 
 * `realism` (`fal/MiniMax-H3-Realism-People-LoRA`) is a *style* LoRA, not a turbo one — realistic people, trigger
   word `r34l1sm`. Same reference tree as larry under a `diffusion_model.` prefix, attention only (`qkv_proj` /
@@ -35,8 +36,8 @@ The two supported LoRAs ship in different layouts:
   (`[96768, 2688]`), so those deltas have no counterpart to fold into.
 
 `H3_LORA` selects the larry file (`off` skips loading it), `H3_LIGHTX=off` skips lightx, `H3_REALISM=off` skips
-realism, `H3_JOYFOX=off` skips joyfox, `H3_LORA_DEFAULT` picks which set starts folded, and `H3_LORA_STRENGTH` is
-the larry card's sharpness/artifact dial.
+realism, `H3_JOYFOX=off` skips joyfox, `H3_LIGHTX8=off` skips the lightx 8-step file, `H3_LORA_DEFAULT` picks which
+set starts folded, and `H3_LORA_STRENGTH` is the larry card's sharpness/artifact dial.
 """
 
 from __future__ import annotations
@@ -51,6 +52,7 @@ LARRY_REPO = os.environ.get("H3_LORA_REPO", "larryvrh/MiniMax-H3-Turbo-Lora")
 LARRY_FILE = os.environ.get("H3_LORA", "minimax_h3_turbo_v4_step600_ema.safetensors")
 LIGHTX_REPO = os.environ.get("H3_LIGHTX_REPO", "lightx2v/Minimax-h3-Turbo")
 LIGHTX_FILE = os.environ.get("H3_LIGHTX_FILE", "minimax_h3_fl2v_turbo_4step_v0.1.safetensors")
+LIGHTX8_FILE = os.environ.get("H3_LIGHTX8_FILE", "minimax_h3_fl2v_turbo_8step_v1.0_bf16.safetensors")
 LIGHTX_ALPHA = 8
 REALISM_REPO = os.environ.get("H3_REALISM_REPO", "fal/MiniMax-H3-Realism-People-LoRA")
 REALISM_FILE = os.environ.get("H3_REALISM_FILE", "h3-realism-people-t2v-i2v-r2v.safetensors")
@@ -107,19 +109,20 @@ def _load_larry(inner_dim: int) -> dict:
     }
 
 
-def _load_lightx() -> dict:
+def _load_lightx(file: str = LIGHTX_FILE) -> dict:
+    """A diffusers-native PEFT checkpoint from `lightx2v/Minimax-h3-Turbo` — keys map name-for-name."""
     from huggingface_hub import hf_hub_download
     from safetensors.torch import load_file
 
-    lora = load_file(hf_hub_download(LIGHTX_REPO, LIGHTX_FILE))
+    lora = load_file(hf_hub_download(LIGHTX_REPO, file))
     suffix_a, suffix_b = ".lora_A.default.weight", ".lora_B.default.weight"
     bases = sorted({key[: -len(suffix_a)] for key in lora if key.endswith(suffix_a)})
     ranks = {lora[f"{name}{suffix_a}"].shape[0] for name in bases}
     if len(ranks) != 1:
-        raise ValueError(f"Mixed LoRA ranks in {LIGHTX_FILE}: {sorted(ranks)}")
+        raise ValueError(f"Mixed LoRA ranks in {file}: {sorted(ranks)}")
     entries = [(f"{name}.weight", lora[f"{name}{suffix_a}"], lora[f"{name}{suffix_b}"]) for name in bases]
     return {
-        "label": f"{LIGHTX_REPO}/{LIGHTX_FILE}",
+        "label": f"{LIGHTX_REPO}/{file}",
         "scale": LIGHTX_ALPHA / ranks.pop(),
         "entries": entries,
     }
@@ -195,6 +198,8 @@ def apply_lora(transformer) -> str | None:
         sets["larry"] = _load_larry(inner_dim)
     if os.environ.get("H3_LIGHTX", "on").lower() not in ("", "off", "none"):
         sets["lightx"] = _load_lightx()
+    if os.environ.get("H3_LIGHTX8", "on").lower() not in ("", "off", "none"):
+        sets["lightx8"] = _load_lightx(LIGHTX8_FILE)
     if os.environ.get("H3_REALISM", "on").lower() not in ("", "off", "none"):
         sets["realism"] = _load_realism(inner_dim)
     if os.environ.get("H3_JOYFOX", "on").lower() not in ("", "off", "none"):