Alice T2V Generator (Eric)

This is the half of the pack that actually does something visible. Feed it the pipeline output from Alice T2V Loader (Eric), type a sentence, and out the other end comes a standard ComfyUI frames image batch you can hand straight to VHS_VideoCombine or SaveAnimatedWEBP. Text-to-video, locally, no API.

Alice T2V is Mirage's open 14B Mixture-of-Experts model, and the MoE trick is the whole story: instead of one giant DiT doing every denoising step, it runs two experts - a high-noise expert that handles early denoising (composition, structure) and a low-noise expert that handles the late detail pass. That's the same architecture play as Wan 2.2, and it's worth knowing the Reddit rumor mill already suspects Alice is a lot closer to a Wan 2.2 derivative than the marketing admits. Either way, treat it as what it is: a brand-new model with no LoRA or ControlNet ecosystem yet, not a drop-in Wan replacement.

How it works

The switchover between experts isn't something you tune - it's baked into the model config as an SNR boundary (boundary=0.875). At the default 40 steps, that works out to roughly 35 steps on the high-noise expert and 5 on the low-noise one.

Now the part that trips everyone up: the README shouts "state-of-the-art in 4 inference steps" (that's score-regularized consistency distillation, rCM - the same distilled-family trick as Lightning or Turbo, but for video). Yet the node's own default is 40 steps, and the tooltip is blunt that below ~20 steps detail quality visibly degrades. Don't chase the marketing number. The distilled speed-up is real if you need quick drafts at 4-20 steps, but the tuned default is the quality sweet spot. And unlike image distillation's CFG-0-2 rule, this model wants real guidance - cfg_high_noise 4.0 and cfg_low_noise 3.0, one CFG per expert.

The generator converts Alice's [C, T, H, W] output in [-1, 1] into ComfyUI's [T, H, W, C] float32 in [0, 1] - a plain IMAGE, so it plugs into anything that takes a frame batch.

The inputs that matter

• prompt - natural language, not a tag list. Include subject, action, setting, lighting, camera style. Max 512 tokens; longer prompts get truncated.
• resolution - pick from the presets: 1280*720 (16:9), 720*1280 (portrait), 832*480, 480*832 (lower-VRAM 16:9/portrait). The 960*960, 1024*576, and 576*1024 options come from the README but aren't in Mirage's official config, so expect them to be untested. custom lets you type dimensions (auto-snapped to multiples of 16).
• frame_num - 81 = 5 seconds at 16 fps. Must satisfy (n-1) % 4 == 0; invalid values are auto-corrected, not an error.
• steps - default 40. The node's own tooltip: fewer than ~20 "noticeably degrades detail quality."
• solver - unipc for ≤40 steps (the default), dpm++ if you go 50+.
• shift - default 12.0. Higher pushes more steps onto the structure expert; lower gives the detail expert more work.
• seed - -1 for a random seed each run.
• negative_prompt (optional) - here's the pro tip: leave it empty. Empty uses Alice's built-in Chinese negative prompt, which was tuned alongside the model and is strongly recommended over anything you'd type.

Wiring and install

The graph is just three nodes: Loader → Generator → VHS_VideoCombine. The install is the pack's shared setup - 27 GB of weights from gomirageai/Alice-T2V-14B-MoE, the Alice source vendored via setup_vendor.py, and one pip install:

huggingface-cli download gomirageai/Alice-T2V-14B-MoE --local-dir "D:/models/Alice-T2V-14B-MoE"
cd ComfyUI/custom_nodes && git clone https://github.com/EricRollei/Eric-Alice-T2V-ComfyUI-Wrapper
cd Eric-Alice-T2V-ComfyUI-Wrapper && python setup_vendor.py   # auto-clones the Alice repo into vendor/
pip install easydict

Restart ComfyUI, drop in the example workflow from the README (workflows/Alice-T2V-Eric-326.json), and set the Loader's ckpt_dir to the folder from step one.

Common issues

• "Pipeline returned None" - the generator is explicit that this happens in a distributed context where rank ≠ 0. If you're not running multi-GPU, it usually means the pipeline got evicted or never loaded; re-run the Loader.
• The "4-step" marketing vs. the 40-step default - the tension is by design. If your output looks soft, you're below ~20 steps.
• Resolution came out slightly different than you typed - it snapped to the nearest multiple of 16 (VAE stride requirement).
• VRAM - a 1280×720, 81-frame clip is a real render. With the Loader's offload_model=True + t5_cpu=True you're in the ~28 GB ballpark; turn both off and you need 70+ GB.

One honest caveat to carry in: this is a very new model with a very thin community footprint, so you're largely on your own with the prompt style. The defaults are well-tuned though - prompt in full sentences, keep the negative empty, run 40 steps, and you'll get a coherent 5-second clip that most people would not guess came off a consumer setup.