David Comfort compares Kling and Seedance in film-reference video pipeline

• David Comfort's best filmic-video route was a six-link pipeline: prompt template, Nano Banana 2 at 2K, 1920 downscale, AuraSR 4K upscale, Seedance 4K, then neutral grain, as laid out in his pipeline post.
• The priced version came to about $6.40 per hero shot, with Seedance 4K at $6.22 for 4 seconds in Comfort's step-by-step breakdown.
• Kling behaved like the camera-control model, while Seedance behaved like the detail model; Comfort's model-routing note put Kling's obedience at 0.045 to 0.093 across registers and Seedance's lit-photo motion excess at 0.188.
• The workflow moved out of prompt boxes and into a local test harness: Comfort said Claude Code plus fal APIs let him run image and video-processing tests on a MacBook Pro.
• Seedance also hit a platform constraint: Comfort's API error post showed a likeness/private-info rejection, and his follow-up image showed the source he described as pretty vanilla.

Comfort's thread reads like a lab notebook: the priced six-step chain includes a $0.02 AuraSR pass and a $6.22 Seedance render, the Fable evaluation screenshot scores camera motion at 0.098 versus 0.040, and the next-test note names the variables he still wanted to isolate: stills, prompting, model choice, post, resolution, and upscaling. The useful phrase is “route by deliverable,” which Comfort's pattern note reduces to Kling obeys, Seedance resolves.

Comfort's step-by-step breakdown assigns each link a job and a price:

1. lit_photograph template: fine-art photographic register, motivated chiaroscuro key, true materials, explicitly not a painting.
2. Nano Banana 2 at 2K: composes the source frame with enough pixels that important details are not ambiguous, about $0.10.
3. Downscale to 1920: avoids an AuraSR failure mode above roughly 1920-wide input, where the 4x output can return a dangling URL.
4. AuraSR to 4K: sharpens existing content without inventing texture, about $0.02.
5. Seedance 4K: carries the motion and preserves fine detail, priced by Comfort at $6.22 for 4 seconds.
6. Neutral grain pass: ffmpeg grain plus halation, with no color curve, to add optical texture while leaving the grade for the scene.

The pipeline is overbuilt in the good way: every step has a killed alternative behind it.

Comfort reduced the workflow to two constraints in his summary: the stills need to be filmic, and the videos need to be obedient.

He also moved the work into a programmable harness. In a reply to fal, Comfort said he ran the experiments with Claude Code and an app for APIs, image processing, and video processing on a MacBook Pro.

The first split was temporal activity, not prettiness. The Fable evaluation said Seedance turned an “extremely slow, almost imperceptible push-in” into a full push from wide to medium, while Kling kept the tavern-wide framing closer to the instruction.

The same evaluation put Seedance at 0.098 mean inter-frame score versus Kling at 0.040, about 2.4x more temporal activity. Tiny benchmark, useful benchmark.

Comfort's model-routing note gave the cleanest split:

• Kling: style-independent obedience, 0.045 to 0.093 across every register.
• Seedance: more camera travel as the style gets more cinematic.
• Lit-photograph register: Seedance reached 0.188, which Comfort called roughly 4x the instruction.
• Routing rule: Kling obeys, Seedance resolves.

That last line is the keeper for creative pipelines: choose the renderer for the failure mode the shot cannot afford.

Comfort also ran film-reference comparisons rather than only synthetic tavern prompts. The evidence set includes a Gone with the Wind source shot, a Kling pass, a Seedance pass, then a second reference pair from Her.

The test design is simple and strong: hold the cinematic target steady, swap the video model, then judge camera obedience and resolved detail separately.

Comfort's next-test note moved the experiment toward Pan's Labyrinth and listed the variables he wanted to isolate:

• the source still
• the prompting
• the video model
• post-processing
• model resolution
• upscaling

That list matches the earlier priority stack in his pattern note: register inheritance first, then prompts, post, and resolution.

Seedance rejected one source image in Comfort's later error post with a JSON message saying the media may contain likenesses of real people or other private information. The same screenshot labeled the failure content_policy_violation with reason partner_validation_failed.

Comfort then described the source as a pretty vanilla image in his reply, while another reply said he was trying older, larger “actors” so the image would not look like anyone.