Where 'no X' is ignored, where it adds X, and what to write instead: why negative prompt lists fail across MiniMax H3, LTX 2.5, Kling, and Google Flow.
OMNI LabsUpdated
Key takeaways
MiniMax H3 possesses no negative prompt field; trailing exclusions directly inject those concepts into the generation.
H3 operates with guidance scale 1.0; seeds govern blocking while camera keywords yield identical pixels (PSNR infinite).
LTX 2.5 runs distilled at cfg 1.0 where negative lists are completely inert; typing 'no text' adds text to empty regions.
Raising LTX cfg to 3.0 slows step time to 91.3 s and pushes shared VRAM to 9.2 GB, masking sparse prompts with thrash.
Prompt engineers must state positive physical conditions, writing 'bare, clean eyelids' instead of 'no eye makeup'.
MiniMax H3 and the positive conditioning trap
MiniMax H3 has no negative-prompt field. Everything typed is positive conditioning, so a trailing 'no blur, no distortion, no extra fingers' introduces those concepts instead of removing them. Our rule is to delete every negative and state the good version: 'bare, clean eyelids' instead of 'no eye makeup'.
Prompt engineering habits developed on legacy image diffusion pipelines become actively destructive when applied to contemporary video architectures. In text-to-image systems, negative token arrays act as secondary repulsive gradients steering latent vectors away from undesirable clusters. However, MiniMax H3 processes all prompt text through a single positive conditioning encoder. When an engineer appends boilerplate negative phrases—such as no blur, no distortion, no extra fingers—the text encoder registers those exact semantic tokens. Instead of filtering out artifacts, the attention map begins actively materializing blurry textures and limb irregularities across the render.
To prevent self-inflicted defects, studio prompt architecture enforces a strict affirmative discipline. Every negative exclusion is deleted and replaced with an explicit physical description of the desired state. For cosmetic and portrait close-ups, specifying bare, clean eyelids reliably eliminates cosmetic pigment without exposing the latent space to makeup tokens.
Guidance scale mechanics and seed-driven camera blocking
H3 runs with classifier-free guidance off (guidance_scale 1.0), so camera and blocking words are not enforced. A test that reversed the walk direction, added a framing lock and changed Pull Out to Truck Left produced a file with zero changed pixels across 124 frames (PSNR infinite).
Understanding classifier-free guidance explains why camera prompt adjustments frequently fail to alter video blocking. MiniMax H3 executes inference with guidance scale disabled at 1.0. At this setting, text embeddings exert gentle semantic steering rather than rigid spatial enforcement. In our diagnostic benchmarks, modifying a complex prompt—inverting actor walking vectors, specifying locked framing, and swapping a pull-out move for a truck-left tracking shot—generated an output bitstream identical to the previous take. Across all 124 frames, zero pixels changed value, resulting in infinite PSNR.
The same H3 test with only the seed changed gave a different film (PSNR 15.1). The seed decides the shot, and words about motion or framing do not. We batch three or four seeds and select, instead of rewriting camera language.
Spatial trajectory in unguided models is determined almost entirely by the initial pseudorandom noise seed. When running identical camera prompts with an altered seed, the benchmark produced a completely distinct composition yielding a PSNR of 15.1. Rather than wasting production hours fine-tuning elaborate camera adjectives that the model ignores, our directors generate batches across three or four seeds and select the composition matching the storyboard.
LTX 2.5 distillation, text hallucinations, and guidance thrashing
LTX 2.5 runs distilled at cfg 1.0, where the negative prompt does nothing. Our own runs carried a negative list the whole time and it never had an effect.
Model distillation eliminates inference latency while fundamentally altering parameter responsiveness. Because LTX 2.5 operates as a distilled diffusion model fixed at cfg 1.0, the negative conditioning branch is computationally inert. During early studio deployments, our automated scripts passed extensive negative prompt arrays on every execution; empirical audits confirmed that these negative strings exerted zero influence on output latents.
On LTX, 'no text' on an empty area adds text. The fix is a positive material identity for the empty space, so it reads as a described surface rather than a gap.
Negative instructions become dangerous when applied to untextured spatial regions. When a prompt specifies 'no text' over a plain studio backdrop or clean architectural wall, the attention mechanism interprets the word 'text' as a focal instruction, rendering garbled typographic glyphs across the surface. To protect blank walls, prompts must assign a positive material identity, describing smooth polished plaster, brushed concrete, or seamless matte paper.
Raising LTX cfg to 3.0 to rescue a thin prompt cost 91.3 seconds per step against 18.6 warm, and pushed shared VRAM to 9.2 GB with thrashing. It masked the real cause, a sparse prompt, instead of fixing it.
Artificially forcing classifier-free guidance on distilled architectures degrades system stability. When engineers elevated cfg to 3.0 in an attempt to compel obedience from a sparse prompt, inference times surged from 18.6 warm seconds per step to 91.3 seconds per step, spilling 9.2 GB into shared system RAM. Forcing higher cfg creates severe hardware thrashing without resolving underlying prompt deficiencies.
Cross-platform negative capabilities: Kling and Seedance
Kling 4.0's official guide says music and subtitles are the only negatives truly worth writing, and it advises phrasing the rest positively, for example 'shown in profile while on the phone' rather than 'do not show her face'.
Commercial platforms maintain highly divergent capabilities regarding negative prompts. In Kling 4.0, official engineering documentation explicitly restricts valuable negative conditioning to suppressing unwanted background audio and burnt-in subtitle streams. For visual staging and actor blocking, Kling's documentation advises affirmative phrasing: directing a subject to appear in profile while using a phone proves vastly more reliable than instructing the generator not to show her face.
Negative conditioning behavior and prompt architecture across video rails · measured 16 Aug 2026
Generation Rail
Negative Support Architecture
Observed Response to Negative Text
Studio Prompting Protocol
MiniMax H3
No negative field (guidance_scale 1.0)
Trailing negative phrases materialize concepts
State positive condition like bare, clean eyelids
LTX 2.5
Distilled cfg 1.0 (negative branch inert)
Exclusion 'no text' adds text to empty areas
Describe positive material identity for surfaces
Kling 4.0
Supported for audio and subtitles only
Broad visual negatives increase error rates
Phrase camera and action affirmatively
ByteDance Seedance 2.5
Dedicated timeline and global negatives
Honours negative constraints accurately
Input targeted negative blocks into timeline slices
Google Omni (Flow)
No negative field
Excluded terms reported appearing in renders
Specify positive locked framing and vehicle geometry
Measured on 2026-08-16; H3 seeds govern blocking while camera wording yields identical pixels.
Seedance 2.5 in our prompts honours negatives: 'no identical clones', 'no distorted facial features or extra fingers' and 'no subtitles, no background music' went into the timeline slice and the global supplement, as the official guide describes.
Unlike distilled models, ByteDance Seedance 2.5 features a dual-branch conditioning engine that actively enforces negative parameters. In our production templates, injecting explicit negative blocks into timeline slices reliably suppresses character cloning, anatomical defects, and unwanted UI overlays.
Google Omni limitations and legacy parameter fallacies
Google's Omni model has no negative field, and a Flow user reported excluded concepts appearing in the output. We write the positive state instead, such as 'static locked-off framing' and 'unchanged vehicle geometry'. This is a field report, not our own measurement.
On Google Flow, Google's Omni model provides no dedicated interface or API field for negative conditioning. When creators attempt to insert negative clauses into positive prompt text, the multimodal decoder frequently manifests the excluded objects directly in the frame. In alignment with field reports, our studio writes affirmative constraints, declaring static locked-off framing and unchanged vehicle geometry to maintain spatial stability.
A Veo-style negativePrompt field is a different mechanism that takes unwanted concepts as a list. Porting that habit onto the H3 or Omni prompt is the mistake the negative section exists to prevent.
Conflating distinct technical architectures causes systemic pipeline failures. In Google Veo, a dedicated negativePrompt argument accepts structured lists of unwanted concepts. Attempting to port Veo-style negative lists into single-stream models like MiniMax H3 or Google's Omni model corrupts visual conditioning, proving that prompt architecture must always conform to model-specific mechanics.
