Image-to-Video Rules: Everything Must Already Be in Frame One
Why image-to-video diffusion models cannot simulate causation: authored first frames, preventing spontaneous materialisation, and scale physics.
OMNI LabsUpdated
Key takeaways
First-frame conditioning acts as a visual hint rather than a mathematical lock in models like Seedance 2.5.
Prompts must reference @Image 1 as the exact first frame and forbid unlisted subjects from spawning.
Physical products must be standing fully formed in frame one to avoid spontaneous emergence from surfaces.
Granular particles and airborne dust frequently collapse into fluid smoke; keep particle bursts short and near source.
MiniMax H3 first-frame generation collapses on non-human subjects unless plates are cropped tightly.
First-frame hint conditioning and hallucinated additions
A Seedance 2.5 first-frame still is a hint, not a lock. On an episode shot on 18 October 2026, naming the other character's position spawned a second person, and re-describing the room made the model rebuild a wider shot.
In generative video production, directors frequently assume that supplying an initial still locks all visual elements into an immutable baseplate. However, diffusion decoders treat reference stills as soft semantic conditioning rather than unyielding geometry. During production on 18 October 2026, when our prompt described where an off-camera conversational partner was positioned, the model erroneously synthesized a second human figure directly into the scene. Furthermore, including superfluous descriptive text about room architecture prompted the model to discard the plate's focal framing and regenerate a wider environment.
The wording that worked for a single-character first-frame shot: 'Use @Image 1 as the exact first frame and continue forward from that moment', then only what moves, plus 'do not add any person or object that is not in @Image 1'.
Controlling first-frame adherence requires rigorous negative constraints. Prompts must explicitly instruct the diffusion model to adopt @Image 1 as the exact starting frame and limit text instructions solely to the mechanical motion of existing subjects. Appending an explicit negative directive forbidding the addition of unlisted entities prevents the latent decoder from populating the background with extraneous figures.
Authoring physical presence to prevent spontaneous emergence
On Flow, an opening on bare wet stone made the model birth the product out of the stone. The fix was to author the opening frame with the object already standing there, under-lit, and to say it is present from the very first frame and never appears or emerges.
Diffusion models lack intuitive understanding of physical object permanence and temporal causation. When prompted with a video opening on an empty surface followed by the instruction that a luxury fragrance bottle appears, the generative model cannot visualize a logical entrance. In our tests on Google Flow, the model simply caused the product to morph upward directly out of the wet stone like an organic growth. To achieve a realistic commercial sequence, the bottle must already be authored standing in position within frame one, using dramatic low-key lighting that brightens over time.
Generative failure modes and first-frame prompt mitigation rules · measured 18 Oct 2026
Video Rail
First-Frame Challenge
Observed Generative Defect
Mandatory Studio Protocol
Seedance 2.5
Extra subjects and room redraws
Model spawns unlisted characters
Lock prompt to @Image 1; forbid additions
Google Flow
Object emergence on blank plates
Products morph out of solid surfaces
Place product in frame one under low light
Omni Flash
Airborne dust and fine glitter
3 of 4 clips render grains as smoke
Short particle bursts localized near source
MiniMax H3
Non-human chrome and prop plates
5 of 5 prompts collapsed after two seconds
Re-crop plate tightly or route to LTX 2.5
Enforce the burst rule so dynamic particles fly in from outside the frame.
A silhouette or lost-in-darkness beat erased the product's surface to a dark cylinder on both Flow and Seedance, and then the graphics had to re-materialise. We drop those beats from product films and vary the light instead.
Cinematic lighting cues that rely on total shadow create catastrophic texture loss in diffusion pipelines. When a camera moves through an unlit silhouette beat, the latent encoder loses track of brand typography, glass reflections, and label engravings. Upon returning to key lighting, the model is forced to hallucinate new surface graphics from scratch, corrupting brand fidelity. Commercial lighting plans must maintain subtle fill lighting across product surfaces at all times.
Particle physics limitations and fluid collapse in Omni Flash
Airborne glitter or dust does not render as particles on Omni Flash. In 3 of 4 clips of a gold-dust prompt the grains came back as curling yellow smoke or a flame-like ribbon. Granular material reads as fluid, so we keep particle shots to short bursts near the source.
Rendering distinct granular dynamics pushes current diffusion attention mechanisms beyond their resolution limits. When generating airborne particles, high-velocity glitter, or cosmetic powders in Omni Flash, individual grain trajectories cannot be resolved across successive latent frames. In 3 of 4 clips evaluated during our gold-dust benchmarks, discrete particulate matter merged into continuous streams of fluid smoke or wavy ribbon-like structures. To maintain photographic realism, particle effects must be staged as brief localized bursts close to the origin point rather than sustained atmosphere.
Omni Flash animates a body in place for as long as asked, but it resists sustained travel across the frame. A dune slide settled into a seated pose and stayed there even with an explicit 'never stop sliding' line.
Spatial translation presents another major generative hurdle. While Omni Flash excels at animating complex gestural movements for stationary subjects, it actively resists sustained linear displacement across broad landscapes. In our desert slope tests, a character instructed to slide continuously down a sand dune decelerated within two seconds, settled into a stationary sitting position, and refused to move further despite emphatic prompt instructions.
Frame-zero scale physics and volumetric ground collisions
Physics at frame-zero scale decided one walking shot. A normal-scale stride of about 1 metre cannot span a 25-metre street, so the feet sank into the gaps. Re-rendering the character seven storeys tall made the gap narrower than the shoe and the walk needed no more wording.
When a generative model attempts to animate locomotion across complex architectural geometry, physical proportions established in the initial plate govern mechanical plausibility. In a conceptual sequence requiring a human subject to traverse broad cobblestone pavers, a standard human stride of about 1 metre proved mathematically incompatible with deep 25-metre roadway gaps, causing the character's footwear to clip deeply into the ground geometry. Rather than laboriously adjusting movement prompts, our team recomposed the initial plate to depict the character seven storeys tall. Because the architectural gap was now narrower than the subject's footwear, the model animated a clean stride across the terrain without any additional prompt modifications.
The burst rule: berries or particles must 'fly in from outside the frame', never spawn inside it.
Whenever dynamic props or secondary particles enter a scene, their trajectory must originate beyond the viewport borders. Instructing particles, food ingredients, or water droplets to appear within the active frame prompts the model to generate them out of thin air, breaking physical realism. Directing objects to enter from outside the frame maintains logical spatial continuity.
MiniMax H3 subject collapse and plate composition remedies
On a chrome balloon concept with no people, the H3 first-and-last-frame model collapsed in 5 of 5 prompts, each time after about two seconds. A static camera with described subject motion held on the same kind of plate in LTX 2.5.
MiniMax H3 demonstrates exceptional facial fidelity when conditioned on human actors, but displays severe structural fragility when presented with abstract non-human props. In a benchmark series testing a metallic chrome balloon plate with no people, the H3 first-and-last-frame model collapsed into chaotic visual noise across 5 of 5 prompts, disintegrating after roughly two seconds of runtime. In contrast, routing the identical plate through LTX 2.5 with a static camera instruction and described subject movement maintained solid geometry throughout the take.
H3 first-frame clips on wide full-body plates changed the subject's scale in our tests. A profile plate is only safe if scale cannot change, so we treat profile plates as still-only, and we verify framing by extracting frames, not by a completed render.
The H3 model reacts to what the frame contains, not to what the text says. Our person-derived rule that large subjects hold did not carry over: a large, sharp, centred object collapsed anyway. Rewritten camera wording changed nothing, and the lever that worked was re-cropping the plate.
