Luma Ray 3: HDR when audio is someone else's job
Luma Ray 3, launched 18 September 2025, makes native 16-bit HDR video at 1080p for grading. On Versely, Luma runs as Ray 3.2 and Ray 2, in standard SDR.
6 min read
MiniMax H3 at 2K when the plate will be graded
2K cinematic with native stereo. Use it when the next step is a grade, not a TikTok upload.
6 min read
Qwen Image 3 Pro when the slide is the product
Dense text and newspaper-like pages. Use it when the still is a slide, not a photograph.
6 min read
Runway Gen-4.5 is control, not a one-click model
Motion brush, camera, world consistency. Weak on native audio and arena Elo. A director's tool, not a volume engine.
6 min read
Seedance I2V is still the still-to-motion job
Even if T2V Elo moves, Seedance's I2V lead is the reason you animate a pack shot here.
6 min read
Seedream 5.0 Pro Edit when the layout is the job
Edit endpoint vs generate. Infographics and dense type are an edit of a layout, not a photoreal hero.
6 min read
Veo 3.1 still wins always-on audio
Eleventh on some Elo boards and still the pick when audio cannot be optional. Official 4K tier. Timestamp-block prompting is Google's multi-shot method.
6 min read
Fast-motion subjects and the models that survive
Subject speed breaks video models before camera speed does. A per-frame overlap test that predicts limb doubling, plus the shot design that buys real headroom.
9 min read
Fluid, fire and smoke: where video models diverge
Pours, flame and volumetric smoke fail in three different ways. How to rank them, and a rule for what to generate, what to composite, and what to cut.
9 min read
What HappyHorse's 38-second render claim means
Alibaba quotes ~38 seconds per 1080p HappyHorse 1.1 clip on one H100. What a hardware-conditioned figure predicts about a hosted queue, and what it can't.
9 min read
Lyria 3.5: when a full song beats a looped bed
Lyria 3.5 shipped into Flow Music on 29 July 2026 with direct tempo and duration control. When a three-minute vocal earns its place, and when a bed wins.
9 min read
On-screen text in generated video: what holds up
Text a video model draws inside the frame has to be right 200 times, not once. The size threshold where letterforms break, and the clean-plate rule.
9 min read
Picking the chat model behind the agent
The chat model plans; the generate model renders. Change chat_model to pick the brain, and pin generate models separately when the clip has to match.
7 min read
Reve 2.1 leads the image editing board
Reve 2.1 holds first on Artificial Analysis image editing at 1262 Elo. What an editing score actually measures, and how to turn a rank into a revision workflow.
8 min read
Reve 2.1 makes 4K without an upscale pass
Reve 2.1 generates natively at 16 megapixels. Where that deletes an upscale step from a print or thumbnail pipeline, and where a dedicated upscaler still wins.
8 min read
Spec-sheet claims that don't survive testing
Six spec numbers that promise more than they deliver: max duration, resolution, reference count, fps, native audio, multi-shot, and one test to run on each.
9 min read
Two-pass regeneration vs native high resolution
Some models reach 2K by regenerating a low-res render; others generate high-res natively. What each does to fine detail, temporal stability and render time.
9 min read
Veo burned in subtitles you never asked for
Quoted text and contractions trigger Veo auto-subtitles. The prompt rewrite that stops it, and why writing 'no subtitles' makes it worse.
9 min read
What turbo tiers actually remove from a model
Fast, lite and mini variants cut sampling depth, resolution, reference slots or options, and each cut leaves its own artifact. Predict which shots survive.
9 min read
Where motion coherence breaks as clips get longer
A 15-second ceiling is rarely 15 usable seconds. The three ways long clips fall apart, a ladder test to find your safe duration, and where to place the cut.
9 min read
Which image models hold architectural perspective
Bowed lines, disagreeing vanishing points and drifting facade counts sink generated architecture. The traits that help, and the prompt that fixes interiors.
9 min read
Which video models hold up with crowds in frame
Background people break more shots than hero subjects. The pixel budget that predicts failure, plus staging tactics for street, stadium and venue shots.
9 min read
Which video models speak more than English
Native audio is a capability flag, not a language list. A one-generation test for any model, and the rule for generating natively versus dubbing after.
8 min read
Why Flux ignores your negative prompt
Flux ignores your negative prompt because the family has no CFG channel. The field is decorative; describe what you want instead.
9 min read