two-stage + dev runs but output is mostly noise (distilled control looks good)

[natestrong]

Hi, thanks for the repo.

I’m seeing a quality issue with non-distilled generation.

Repro

PROMPT="A cinematic shot of a red fox walking through a snowy forest at sunrise, realistic motion, natural lighting"
SEED=1543716980

Distilled control (looks coherent):

uv run python scripts/generate.py "$PROMPT" \
  --pipeline distilled --model-variant distilled \
  --weights weights/ltx-2/ltx-2-19b-distilled.safetensors \
  --gemma-path weights/gemma-3-12b \
  --height 320 --width 512 --frames 9 \
  --steps 8 --seed "$SEED" \
  --output outputs/distilled_control.mp4

Two-stage dev (valid mp4, but mostly noisy):

uv run python scripts/generate.py "$PROMPT" \
  --pipeline two-stage --model-variant dev \
  --weights weights/ltx-2/ltx-2-19b-dev.safetensors \
  --distilled-lora weights/ltx-2/ltx-2-19b-distilled-lora-384.safetensors \
  --distilled-lora-scale 0.8 \
  --spatial-upscaler-weights weights/ltx-2/ltx-2-spatial-upscaler-x2-1.0.safetensors \
  --gemma-path weights/gemma-3-12b \
  --height 320 --width 512 --frames 9 \
  --steps-stage1 15 --steps-stage2 3 \
  --cfg 5.0 --cfg-stage1 5.0 --guidance-rescale 0.7 \
  --seed "$SEED" --low-memory \
  --output outputs/two_stage_dev.mp4

Result
Distilled: coherent video.
Two-stage dev: has rough scene structure, but remains heavily noisy/grainy.
I also tried changing stage1 steps (8, 15, 30), and the noise issue persisted.

Can you confirm whether this is expected currently, or if there are recommended settings/known fixes for two-stage dev quality?