Geiru Context Suite keeps long MiniMax H3 chains from drifting
The free ComfyUI node pack, released November 12, holds one clean frame from the start of a chain in front of every later shot.

Key takeaways
- Geiru Context Suite is a free ComfyUI node pack for long MiniMax H3 video, published to the ComfyUI registry as version 1.0.0 on November 12, 2026, under the GPL-3.0 license.
- In the author's own 10-clip test, color drift from clip 1 fell from 6.4 with no Keystone to 3.9 with the keystone and 2.9 with the keystone plus Late Lock.
- On a 50-clip chain of about 5 minutes 20 seconds, the unprotected version had fallen apart by clip 20 while the protected one stayed clean but somewhat darker than clip 1.
- For chains that carry sound, clip lengths should be 141, 192, 243, 294 or 345 frames, and the pack needs one H3 Motion Context pack installed, not both.
Geiru Context Suite, a free ComfyUI node pack from the developer roycho87, keeps one clean frame from the start of a MiniMax H3 chain in front of every later shot, so a long video stops losing quality clip by clip. The pack's pyproject.toml lists version 1.0.0 and the GPL-3.0 license, and the ComfyUI Wiki dates its ComfyUI registry publication to November 12, 2026.
The problem it answers is built into how H3 chains work. H3 renders a few seconds in one pass, so anything longer is a chain of clips, and each new clip starts from the tail of the one before. You tell the model where to start, never where to end. Detail is lost a little in every hand-off, and because each clip begins from the one before it, the loss stacks up over the chain, as the pack's README puts it.
How Geiru Keystone holds the chain steady
Clip 1's first frame is the only frame in a chain made before any drift, and the author's testing found that the first frame is what H3 leans on most. Geiru Keystone takes that frame, the keystone, and locks it into the start of every later shot. It can also tack a frame from the end of clip 1, the capstone, onto the back of each shot. In the pack's 192-frame example, the capstone is frame 187.
A capstone locked in from the first sampling step pulls a shot's pose back toward clip 1, which fights the motion. Geiru Late Lock answers that by choosing the step at which the keystone and capstone latents come in: the shot settles on its own pose first, then the reference frame is swapped in for the rest of the run. The default swap_step is 4, and the README suggests the last quarter of your step count, which is 6 on an 8-step run.
With Geiru Context in front of Keystone, the held frame is a real frame from the previous shot and it is put back after sampling. The saved clips keep their own first frame, and nothing from clip 1 ends up in the video.
One switch turns the protection on and off
Three nodes do the work. Geiru Keystone locks the reference frame in. Geiru Late Lock times it. Geiru Context Noise, an idea the README credits to the developer ethanfel, gives the held frames part of the sampler's current noise, fading to none at the join, so H3 does not copy them so exactly.
Geiru Context's degradation_protection switch turns all three on or off together. Set them up once, then flip the one switch: on for a long stretch of one continuous scene, off around cuts and scene changes. In the author's testing, a hard cut or a new environment already resets the drift on its own, so protection is only needed while the scene stays the same. When the scene does change, point Keystone's source_clip at the first clip that fully shows the new scene, not the transition clip.
The wiring runs Geiru Context into Geiru Context Noise, then optionally Geiru Anchoring, then Geiru Keystone, then Geiru Late Lock, then the sampler, with the model plugged through Context Noise, Keystone and Late Lock.
The settings that work
The README's recommended settings are Geiru Context head_frames 39 with drop_last 17, Geiru Anchoring tail_frames 56 with drop_first 17, and Late Lock swap_step 4. Anchoring's frame count has to outnumber Context's. Context noise tested best at strength 1, which is now the pack's default.
Two frame rules trip people up. Clip lengths have to be 17n + 5 frames, the series 5, 22, 39, 56, 73 and on to 192, and the drop settings move in steps of 17. For chains that carry sound, use a clip length that is a multiple of 3: 141 frames (5.9 seconds), 192 (8 seconds), 243 (10.1 seconds), 294 (12.25 seconds) or 345 (14.4 seconds). H3's audio runs at 40 steps a second against 24 video frames, so only every third frame lands on an audio step. At other lengths the held audio is cut a third of a step off from where it belongs, and the sound builds up and gets louder down the chain.
Because each clip is saved whole, with its overlap included, you trim when you join. The pack's sane_clown_combiner.json workflow cuts the last 39 frames off every clip except the last one and crossfades the audio at each seam.
What the author's tests show, and what they do not
The numbers behind the pack are roycho87's own, not an independent test. In a 10-clip chain, measured about 20 frames from the end of clip 10, color drift from clip 1 fell from 6.4 with no Keystone to 3.9 with the keystone in front, and to 2.9 with the keystone plus the Late Lock capstone. Over-sharpening dropped too. On a 50-clip chain, about 5 minutes 20 seconds across four scenes and rendered twice on the same seed, the unprotected version had fallen apart by clip 20 while the protected one stayed clean, though somewhat darker than clip 1.
One caveat comes from the author. Which speed LoRA you use matters, because a LoRA changes how the model reacts to the partly-noised held frames. All of the testing used one 8-step turbo LoRA, and a tester still saw degradation with a different turbo even with context noise at 1.
Where it still drifts
Protection is not absolute. On a fully locked-off static shot the texture holds but the color can still slowly darken, a limit the author notes. The 50-clip test also stayed somewhat darker than clip 1 even where it held, so a chain that needs exact color matching may still need a grade afterward.
Hybrid Windows: two minutes in one render
Not every long video is a chain. Geiru Hybrid Windows renders the whole thing in overlapping windows in one pass, so memory stays at the size of one window, and it repeats the keystone and capstone in each window. The example uses a context_length of 124 frames with a 39-frame overlap, about five seconds a window. The node builds on kijai's MiniMax H3 context windows in ComfyUI, merged as pull request 17425, and works the way JalenBrunson's ComfyUI-HybridWindows does.
Geiru Window Director is an optional node that goes after it and builds the render one window at a time, with Approve, Re-roll, Rewind and Blend all buttons, so a bad window costs a window instead of the whole video. The author's experiments include a 2 minute 6 second roller coaster in one render, 3,031 frames at 960 by 544 in 16 windows of 226 frames, and a talking portrait at 544 by 960, also 2 minutes 6 seconds. Hybrid Windows is not as good as the continuation nodes at preventing degradation. The author says its job is fast motion that has to be seamless. In the full-step test the background darkened slowly, about 4% over the seated windows and 12 to 18% over the standing ones by window 15.
What you need to run it
Geiru Context Suite needs a ComfyUI build with MiniMax H3 support, and exactly one H3 Motion Context pack: NikoDemon80's ComfyUI-H3-Motion-Context or seitanism's ComfyUI-H3-Motion-Context-MultiRef, not both, because the two packs use the same node names. The suite itself adds no Python packages and no model files, and the install is a git clone into ComfyUI/custom_nodes followed by a restart. ComfyUI Wiki notes the pack is also installable from the ComfyUI registry as geiru-context-suite.
The suite ships nine nodes: Geiru Context, Anchoring, Shot Counter, Chain, Keystone, Late Lock, Context Noise, Hybrid Windows and Window Director, with example workflows for each. The repo carried more than 70 stars, read November 15, 2026.
Clone the repo into ComfyUI/custom_nodes, restart ComfyUI, and start from example_workflows/mad_clown_inventor_example.json, which wires Context, Keystone and Late Lock together.
Sources
- github.com - the repo README: the problem, the nodes, the settings, the author's test figures, install
- raw.githubusercontent.com - version 1.0.0 and the GPL-3.0 license field
- github.com - the author's dated test notes behind the drift figures
- comfyui-wiki.com - registry publication date, node count, install
- freakyjolly.com - a guide; frame-count rules and the recommended settings
- github.com - kijai's H3 context windows that Geiru Hybrid Windows builds on