MiniMax H3 CLIP-Cached Ref2VA (Dual Resolution)

MiniMaxH3CLIPCachedRef2VADualRes is what you get when you combine the two ideas the pack sells separately: the cached Ref2VA reference-to-video path, and the Dual Resolution trick of preparing a base generation and its upscale-target conditioning from one shared set of inputs. If your workflow conditions H3 on reference images or video and continues through a latent-upscale branch, this is the node that keeps the prompt, references, encoder choice, and length in a single place instead of duplicating a whole reference conditioning node for the upscale pass.

Like its FL2VA sibling, this is not an upscaler. It prepares conditioning for two resolutions; a separate latent-upscale node resizes the base latent and hands it to the upscale branch. The node returns three outputs - positive and latent for the base width/height, and positive_upscale (CONDITIONING) for width_upscale/height_upscale. There's deliberately no second latent output: an upscale workflow takes the base latent, resizes it externally, and combines that with positive_upscale, so exposing a second empty latent was pointless and the author removed it.

Inputs that matter

Everything from the single-resolution cached Ref2VA node carries over: clip_name, vae plus audio_vae, prompt, width/height, length (124 ≈ 5 s at 24 fps), and ref_image_size (match scales references to the generation pixel area; max keeps the 2048-px short-edge path for identity fidelity at the cost of slower sampling). The reference slots are the same fixed set - ref_image_0–ref_image_8, ref_video_0–ref_video_2 with matching soundtracks, and ref_audio_0–ref_audio_2 - and the same slot rules apply: unconnected slots are dropped, order is preserved, and MiniMax H3 prompt references like <Picture 1> address the positional order after compaction, not the socket number you wired.

The two new controls are the resolution pair width_upscale/height_upscale and generate_upscale_cond. That boolean defaults to on, and it's the only way to skip the second-resolution encode: turn it off and positive_upscale comes back as None. This is more load-bearing than it looks because the node is one atomic call that returns all three outputs together - bypassing the downstream upscale consumer does not skip the encode on its own. cache_mode (auto/refresh) applies to both resolutions; with auto, if the two resolutions end up presenting the encoder with the same effective reference pixels, they share one cache entry and Qwen3-VL loads at most once per run.

Installing

The pack-wide routine applies: ComfyUI Manager (search "MiniMax H3 CLIP-Cached"), or

cd ComfyUI/custom_nodes
git clone https://github.com/Mu5hr00moO/ComfyUI-MiniMaxH3-CLIPCached

Restart, and the node appears under model/conditioning/minimax/cached. It needs ComfyUI v0.30.0+ with native H3 nodes (validated against v0.34.2), a MiniMax H3 encoder checkpoint in models/text_encoders, and no dependencies beyond the safetensors ComfyUI already bundles.

The honest caveat

Two caveats stack here, so read them both. First, this is a Ref2VA-based node, and the author's own docs flag the Ref2VA path as less battle-tested than FL2VA - it delegates to ComfyUI's stock Reference to Video implementation and shares its cache architecture, but hasn't seen the same depth of real-world workflow testing. Second, this is the least common configuration in the pack: you're combining the newer reference path with the two-resolution bookkeeping, so you're the most likely person to hit something the author hasn't. If a workflow-specific issue shows up, that's a report-upstream moment, not a sign you misconfigured it. And if you don't actually have an upscale branch, skip the Dual Resolution node entirely - the plain cached Ref2VA is simpler and won't tempt you into paying for an encode you don't use.