The 5 open-source AI video tools that shipped, ranked by impact

Five open-source video projects shipped release pages in the ten days through November 3, 2026, and each changes a different part of a working ComfyUI or open-weights pipeline. The five cover a tiled upscaler that pushes MiniMax H3 clips to 8K, an agent harness that generates video indefinitely, a character-reference adapter, a hair-card generator for 3D avatars, and a recipe that fits a 14B model on a 16 GB card.

The criterion, stated in full: every entry shipped in the last ten days, on or after 24 October 2026; each entry's own release page was opened and checked for this piece rather than a roundup's account of it; and the entries are ranked by how much of a working ComfyUI or open-weights pipeline each one changes, with each maker's own stated figure quoted for its entry. Every date and number below was read as of 3 November 2026.

No hands-on tests were run for this list. The ranking weighs documented specs, license terms, hardware needs and price, so the verdicts judge fit, not output quality. One candidate a reader might expect is absent: a ProRes 4444-with-alpha codec for ComfyUI's stock Save Video nodes is not here, because the ComfyUI changelog does not carry it. Five tools make the cut.

5. Wan 2.2 on 8 to 16 GB (Alibaba)

Wan 2.2 is Alibaba's open-source video model, and by the FreakyJolly guide published on November 2, 2026, it is the most-downloaded open video model on Hugging Face. What shipped in this entry is the guide itself, a step-by-step recipe for running the 14B image-to-video or text-to-video experts on an 8 to 16 GB card, where the stock setup wants more than 16 GB and 20-plus steps across two experts.

The recipe stacks GGUF quantization with distilled 4-step LoRAs: the 14B experts as GGUF Q4_K_M at 9.65 GB each, plus LightX2V 4-step LoRAs, for four total steps at CFG 1. Below 12 GB it falls back to the TI2V 5B model or the 14B at Q3. It needs ComfyUI core plus ComfyUI-GGUF, and about 32 GB or more of system RAM so ComfyUI can offload the idle expert. Hugging Face's API showed the Comfy-Org/Wan-2.2 repository at 9,489,939 downloads when read on November 3, 2026. Wan 2.2 is Apache-2.0, so commercial use is allowed, though each LoRA carries its own license. This is the least pipeline-changing entry, because none of the software is new, and the one most readers can run tonight. Where HairDeck below adds a capability, this one removes a hardware barrier.

4. HairDeck (hyz317)

HairDeck turns one character image into stylized hair cards, the ribbon primitives that stylized games and VTuber avatars ship, and generates the spring-bone chains that animate them. It then exports GLB or VRM. The code and weights appeared on GitHub and Hugging Face on October 27 and 29, 2026, according to its ComfyUI Wiki release page. It is Apache-2.0 and ships four checkpoints and 100 test cases with their original card geometry.

On CHARM's own test set, the authors report 44% lower Chamfer distance and more than half the card-orientation error while running 33 times faster. Inference needs Linux x86-64, Python 3.12 and one NVIDIA CUDA GPU, and the VRM export drives Blender 4.2, so this is not a ComfyUI node. A community ComfyUI pack, FengLairh/ComfyUI-HairDeck, appeared on November 1 and is not validated by the release page. HairDeck changes a 3D asset pipeline rather than a video one, which is why it sits below the reference adapter above it.

3. SnapRef v1a (0ppxnhximxr)

SnapRef v1a is a reference adapter for Anima Base v1.0. One image of a character goes in, the prompt describes only the scene, and the adapter holds the face, hair, outfit and rendering style. It works by encoding the reference once, caching what each layer produces, and re-reading that cache at every sampling step, so the reference is never re-encoded. The ComfyUI Wiki page carries a last-updated stamp of October 29, 2026, and the node was updated on October 31; the page's news URL is dated November 3.

The adapter is 190.8M parameters, 382 MB in bf16 and 195 MB as FP8. The model card puts the cost at about 15% over a reference-free generation. The repository's own timing table, measured on an A100 at 832x1216 and 28 steps, has SnapRef at 8.95 seconds per image against 14.36 for ElectronGirl v2, 12.53 for Role-Ref-KV and 7.72 for Anima Base with no reference. The model card lists the failure modes: small details such as earrings and scars can be simplified, a dark-skinned reference can come out lighter, and a scene can override the reference outfit. It ships ComfyUI nodes and five example workflows.

2. HyperFlow Infinity (Video Rebirth)

HyperFlow Infinity is Video Rebirth's agent harness for open-ended video. Give it a story seed and it recursively generates a segment, watches it, folds what changed into a scene graph, writes the next prompt and keeps going. It adds no model of its own: the DiT generates, a vision-language model watches and directs, and each segment builds on the last seven latents of the one before. The repository opened with a single initial commit on October 24, 2026.

The README's own numbers are steep. On one eight-GPU H200 node at 1344x768, a five-second segment totals 49.2 seconds, 32.2 of them in the H3-HyperFlow generator; without FlashAttention-3 the generator alone takes 34.1 seconds per segment. It needs at least four NVIDIA GPUs, or three with OpenRouter for the vision-language side, and only H200 141 GB has been tested. The repository's code is Apache-2.0, but the LoRA weights it drives fall under the MiniMax H3 Community License, which excludes the EU, the UK, South Korea and the US. Video Rebirth reports a one-hour 1344x768 film generated in the loop. It is the biggest conceptual change on this list and the least installable.

1. Tessera (hashu786)

Tessera is a nine-node ComfyUI pack that refines an already-upscaled MiniMax H3 plate tile by tile into 2K, 4K or 8K. It was published on October 30, 2026, by hashu786, and it is the one entry that directly changes what a working ComfyUI pipeline produces. The mechanism is streaming: the pack encodes and decodes through temporal windows so the full-resolution clip never has to sit in memory.

The author measured why tiles should be anchored to the plate rather than to each other. Conditioning each tile on its finished neighbor drifted +1.5, -2.1, -9.3 and -15.5 px across four tiles, so each tile is conditioned on its own crop of the source instead. On the author's reference run, a 15-second 544p clip reached 3840x2176 in three hours on one GPU, with a peak encode-side buffer of about 5.2 GB against 57.6 GB for the whole-clip copy that ran out of memory. Tessera is MIT-licensed and installs by cloning the repository into custom_nodes; it is not yet on the Comfy Registry, and the author's low-VRAM recipe is still to come. Where HyperFlow Infinity needs a rack of H200s, Tessera runs on one card.