DreamX Causal Refiner Loader

DreamX-Creator's second stage is a 5B autoregressive refiner - SR-DiT, one-step, 2x - and DreamX Causal Refiner Loader is how you get it on the card. Four inputs, one output. Two of those inputs are the reason this node gets its own article: they exist to keep a 24 GB Windows machine from quietly dying.

Inputs

model_root (default auto, same search order as everything else in the pack) and dtype (bfloat16 default). Then:

• window_chunk, default 1, range 1–64. The tooltip calls it the "24 GB Windows-safe attention batch" and says to raise it only with measured VRAM headroom. It sets the chunk size for the refiner's windowed attention, both on the model and on every module that flagged use_window_attn.
• kv_history_frames, default 3, range 1–64. How many latent history frames the causal KV cache keeps. Three is the 24 GB-safe preset; nine is the model's own high-memory preset, and the pack is explicit that you select it only with measured VRAM and system commit headroom.

Why both default low: on Windows, WDDM doesn't raise a catchable OOM when it runs out - it spills CUDA allocations into shared system RAM and your run turns into a slideshow instead of an error. Defaulting to the safe values means the pack fails softly on the machine that would otherwise fail confusingly. If you're on Linux with 48 GB of VRAM, taking window_chunk and kv_history_frames up is a legitimate upgrade and the original released preset is 9.

Output

refiner, of the custom type DREAMX_REFINER. It only plugs into DreamX Causal Refine Video, which also wants a Wan 2.2 VAE, positive conditioning from UMT5, and the image sequence to refine.

What gets loaded

Two configs and two checkpoints. The SR-DiT 5B (refiner/sr_dit_5b.pt) plus the Flash latent upsampler (refiner/latent_upsampler_flash.pt), which is the default path - there's an optional causal 2D variant and an optional distilled fast decoder in the bundle, but they're off unless you go out of your way. The loader builds the pipeline at your dtype by temporarily setting torch's default dtype around construction, then hands the whole thing to ComfyUI as a CoreModelPatcher so it can be offloaded between runs.

Worth knowing: the refiner's timestep_shift comes from its own config (5.0), not from a widget, so DreamX AV Flow Shifts is not part of the refiner graph. Don't wire one in.

Install

ComfyUI Manager → DreamX Creator T8, or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-DreamX-Creator-T8.git
cd Comfyui-DreamX-Creator-T8
python -m pip install -r requirements.txt

The refiner needs its two checkpoints and wan2.2_ti2v_5b/, because it uses the Wan VAE for its own encode/decode:

python -m pip install -U huggingface_hub
hf download t8star/DreamX-Creator-Comfy --local-dir ComfyUI/models/dreamx_creator
python scripts\verify_models.py

Common issues

Slow, stuttery refinement with no error. Classic WDDM spill. Drop kv_history_frames to 3 and window_chunk to 1 and re-measure. If you're already there, your input clip is too long or too big for the card - the refiner tiles internally and caps output at 8 megapixels, but a 4K sequence will still lose.

Hash mismatch on sr_dit_5b.pt. These are pickle-backed official checkpoints, and the pack verifies them against model_manifest.json before any load. A partial download is refused rather than half-loaded.

Everything OOMs the moment you refine. The 7B Creator and the 5B refiner are not supposed to be resident together. The pack ships two separate workflows - dreamx_creator_ui.json and dreamx_refiner_ui.json - on purpose. Generate, inspect, then rerun in a second graph so ComfyUI can unload the first model. Refining a clip you already like is much cheaper than refining every seed.

You're expecting a real upscaler. This is a learned 2x refiner on a causal video DiT, closer to a detail pass than to SeedVR2 or a tile-based photo upscale (upscaling.md for the general landscape). It will not invent structure that isn't there.