IAMCCS Cine Reference Latent Control

Character drift is the unsolved problem of AI video - the KB's own character-consistency essay calls it the most-cited unsolved problem in the field. The practical 2026 answer is blunt: stop re-inventing the identity, condition on references. IAMCCS Cine Ref Latent Control is that idea taken to the latent level. It takes up to eight reference images, encodes them with your VAE, and splices them into the video latent at the frame positions your plan specifies, each with its own strength. The model literally sees your character's reference mid-generation instead of trying to remember them from the first frame.

It's the "reference" counterpart to the guide/keyframe insertion the Cine LTX Sequencers do - same family of mechanism, purpose-built for identity anchoring rather than timeline staging.

How it works

Required inputs:

• vae - the video VAE that encodes references.
• latent - the video latent being built.
• plan_json - the plan that decides where references go (the Cine planners' format with per-entry frame data and reference indices).
• default_strength (0.8) - how hard each reference anchors when the plan doesn't specify.
• lock_slots (1) - how many leading reference slots stay locked to the plan's first entries.
• soft_transition_slots (0) - how many references get gentle transition handling instead of hard lock.
• reference_fallback (cycle/clamp/skip) - what happens when the plan asks for a reference you didn't connect.

Optional reference_1…reference_8 are your images. The node encodes each reference, upscales it to the latent's spatial size, and repeats it across the batch if needed so it lines up with the video latent's shape - then writes it in at the planned position with a strength-weighted blend. Outputs: latent, applied_count (how many references actually landed - your best sanity check), and report.

The inputs that matter most

• default_strength - the single most impactful knob. Too high (≥0.95) and the video can lock onto the reference like a slideshow; too low and the identity drifts back. 0.7–0.85 is the sane band.
• lock_slots - raise it if the first N references are your guaranteed identity anchors (character face, costume) and should always apply.
• reference_fallback - skip means "if the plan wants ref 5 and it's not wired, drop it." cycle (default) means "rotate through what I have," which is safer for never-crashing but can put the wrong face in.

Install

In IAMCCS-nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/IAMCCS/IAMCCS-nodes.git

or Manager → search IAMCCS. README floor: ComfyUI ≥ 0.3.0, Python ≥ 3.12, Torch ≥ 2.8. Needs the LTX video VAE and the standard LTX sampling chain around it.

Where people get burned

• Too many references fight each other. Every encoded reference is another constraint. Eight strong references across one shot is a recipe for the model compromising in weird ways. Start with 1–3, add only where drift actually shows up.
• Reference size ≠ latent size. The node upscales references to the latent's spatial dimensions, which means low-res source images get soft anchors. Feed it clean, full-res references at your generation aspect ratio.
• applied_count is your friend. If it's lower than you expect, either the plan asked for unconnected slots (fallback kicked in) or the plan was empty and it fell back to defaults. Check it before blaming visual drift on the model.
• It needs a real plan. Like most of the Cine family, this node's value depends on a plan_json from the planners. With {} it does almost nothing and prints a report telling you so.
• Zero search impressions, one-author tooling. This is squarely an identity-continuity tool - if character consistency across a single long shot is your problem, this is the right tool; if you're trying to keep identity across separate shots, the whole shotboard pipeline (with this node per shot) is the intended system.