GenBench Drops a 5-Model Character Benchmark, and the Winner Isn't Who You Think
GenBench just published its biggest image-model benchmark yet: 5 fictional characters, 5 models, 111 completed images across 150 planned cells. The test wasn't about pretty portraits. It was about whether a model can hold an impossible character brief... a thumb-shaped head, a chin like a shelf, a towering goth silhouette... and keep that signature feature legible at thumbnail size through multiple edit rounds.
The models: GPT Image 2.5 Sunburst (OpenAI, current Genflick Image default), Qwen Image 3, Seedream 5.0 Pro, Grok Imagine Image 2.0, and Krea 2 Large. All run through fal on documented endpoints.
Completion rates tell the first story. Qwen Image 3 completed 29 of 30 supported cells. Seedream 5.0 Pro completed 28. Grok Imagine Image 2.0 finished 24. Krea managed 11 of 15 text-only cells (no edit endpoint). Sunburst, the current Genflick default, completed 19 of 30.
But completion isn't the whole picture. The benchmark found that reference-image edits push unusual anatomy further than stronger text prompts. When Buck Bedrock's cowboy chin grew through two edit rounds on Sunburst, it ended up as an enormous anvil-like jaw. The strongest text-only prompt produced a much more conventional cowboy. That's a workflow insight, not just a model score: if you want extreme character features, iterate on the image, don't just rewrite the prompt.
Genflick's GenBench qualification decision: Qwen first, Seedream second for production. Neither new endpoint is enabled in the product yet... but the benchmark gives the lineup. Read the full benchmark with every image and prompt.
GPT Image 2.5: Sunburst vs Flare, and Why It Matters for Your Keyframes
OpenAI shipped GPT Image 2.5 on September 8, splitting the old single model into two: Flare (faster, lower latency) and Sunburst (premium visual workflows, tighter edit control). Sunburst is the current default in Genflick's Image / Keyframe catalog.
The GenBench benchmark put Sunburst through 30 cells of character exaggeration. It produced the cleanest photographic texture of any model in the test... real skin, individual hairs, specular flash reflections. But it also had the most provider refusals on the curvier female-character briefs. If your character design pushes past conventional anatomy, expect some cells to come back blocked.
The workflow takeaway: use Sunburst for keyframes where texture and edit precision matter most... hero shots, character close-ups, scenes where the image IS the reference for a video generation pass. Use Flare for rapid iteration and storyboards where you're burning through dozens of frames before picking the keepers.
Grok 4.7 Joins the AI Director Lineup
xAI launched Grok 4.7 on September 21... positioned as their most capable model for coding and knowledge work, twice as fast at half the price of comparable models. In Genflick's catalog, Grok 4.7 sits in the AI Director phase... the layer that writes your plan, casts your characters, draws your boards, and structures your scene beats before a single frame of video gets generated.
What does a smarter AI Director mean in practice? Longer planning chains, better instruction following, and... if the GenBench benchmark pattern holds... fewer hallucinated scene descriptions. The AI Director is the first phase of Genflick's three-phase pipeline: Director plans, Image generates keyframes, Video animates them. A stronger director means fewer wasted generations downstream.
Runway Ships a DaVinci Resolve Plugin
Industry context, not a Genflick feature: Runway shipped a DaVinci Resolve plugin on September 23 that lets editors generate images and video, restyle the clip under the playhead with Aleph 2.0, and drop results into the Media Pool... all without leaving the timeline. It's the clearest sign yet that AI generation is moving from standalone web tools into the NLE where post already happens. The plugin is free for DaVinci Resolve Studio users on existing Runway plan credits.
What This Means for Your Next Film
The GenBench benchmark, GPT Image 2.5's split, and Grok 4.7's arrival all point to the same thing: the three-phase pipeline is getting sharper at every stage. Your AI Director plans better. Your keyframe model holds character features through edits. Your video model... Seedance 2.5, which can already carry 50 references through a 30-second clip... has better source material to work from.
Run it all on Genflick. Start with a sentence. Your director takes it from there.
Genflick is building the AI filmmaking stack. Start free, no card. Sign up here.