
FastFrames Video Generation vs Luma Ray 3.2: A Frame Stream Against a Graded File
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 89 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-079 tok/s
- OrcaNEWOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 88 tok/s
- OrcaNEWOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaNEWOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 87 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 121 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 68 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 91 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 44 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 195 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 118 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 187 tok/s
- OrcaOrca: OrcaCyber Zero 1.52026-10-10$3.00 / $7.50 per 1M tokens · 143 tok/s
Luma Ray 3.2 and FastFrames Video Generation are both diffusion video models, both released in 2026, both shipping no earlier than this year's second half — and they are not substitutes for each other in any sense a producer would recognise. Ray 3.2, out since 9 June, terminates in a file: up to 1080p, five or ten seconds, optionally 10-, 12- or 16-bit HDR in the ACES2065-1 EXR standard, which is the format a colourist grades in. FastFrames, published by Apple with Cornell on 16 November, terminates in a stream: 832×480, 81 frames, one frame every 36 milliseconds on a single B200, silent, text-prompt only, weights on Hugging Face. One is built so its output survives a finishing pipeline. The other is built so its output arrives before you have finished watching the last one. Everything below follows from that split, including the uncomfortable fact that neither model has an independent score to settle the question with.
The output format is the argument, not the spec sheet
Start with what each model hands you, because this is where the comparison actually lives.
• Ray 3.2 delivers MP4 by default, and on request delivers 10-, 12- or 16-bit HDR frames in ACES2065-1 EXR — but only at 720p and 1080p, and only for five-second clips. Seamless looping and the EXR path are both five-second features.
• FastFrames delivers an MP4 of 81 frames at 832×480 and 16 frames per second, which is 5.06 seconds of silent footage. There is no EXR path, no bit-depth choice, and no export dialog. The pixel decoder emits 8-bit RGB per token and that is the end of it.
• Resolution is a 1080p ceiling against a 480p fixed point — but the comparison is unfair in both directions, because Ray 3.2's 1080p HDR mode does not exist for ten-second clips, and FastFrames' 480p is not a setting you can raise.
• Control surface: Ray 3.2 takes text, a start frame, an end frame or an existing video to modify, plus up to 16 keyframes per clip and facial performance tracking across up to eight faces. FastFrames takes a text prompt and a seed.
Read that list again and notice what it implies about the buyer. Ray 3.2's feature set is the feature set of a shot that already exists in an edit and needs to be finished, revised or extended. FastFrames' feature set is the feature set of a system that has not decided what the shot is yet.
Ray 3.2, read as a post-production tool
The EXR path is the whole reason this model exists in its current form, and it comes with constraints that most write-ups skip.
It generates at 540p, 720p and 1080p, in five- or ten-second clips, across six aspect ratios covering vertical, square and widescreen — but ten-second generation is text-to-video only, so the frame-conditioned and video-conditioned paths are capped at five seconds. HDR is a 720p-and-1080p, five-second, opt-in setting: you enable HDR to get the EXR export rather than getting it automatically. Audio is the other constraint, and it cuts the way people do not expect. Ray 3.2 does not generate synchronised audio for text-to-video or image-to-video; audio survives only when you are modifying or reframing existing footage that already carries a track. For a model positioned at finishing, that is coherent — the sound comes from the edit, not the generator — but anyone shopping for a one-call video-with-sound model should stop reading here.
The commercial structure matters too. Ray 3.2 is the first model in the Ray family with a full developer API, split into a pay-as-you-go Build tier and a reserved-throughput Scale tier with an SLA. Earlier Ray models were subscription-only, which means the family's pricing history is not a guide to this one's. Per-clip cost moves with resolution, dynamic range and duration together; third-party listings put a five-second 720p clip near $0.30 and a 1080p clip near $1.20, and HDR or EXR export multiplies the clip cost rather than adding to it. Treat those as aggregator arithmetic against Luma's rate card rather than as Luma's own published number.
One naming trap worth the sentence it costs: Luma's version numbers do not ascend the way they read. Ray3.14 shipped on 26 January 2026 and Ray 3.2 on 9 June 2026, so the lower number is the newer model, and Luma has retired the Dream Machine branding that most older guides are still built around.
FastFrames, read as a render loop
FastFrames is not a smaller Ray 3.2. It is a distillation result published as weights, and the architecture explains both the speed and the missing features.
Both checkpoints are autoregressive video diffusion models built on Wan2.1 — FastFrames-S from Wan2.1-T2V-1.3B and FastFrames-L from Wan2.1-T2V-14B — put through a recipe the authors call Causal Forcing++. The paper's contribution is not a faster transformer. It is the deletion of two expensive components around the transformer: the VAE decoder, which cost roughly 147 milliseconds a frame in the baseline and is replaced by a 63.87 million-parameter pixel decoder that lands near 1 millisecond, and the KV-cache update, which becomes a small refinement network costing 2 milliseconds on the small model and 6 on the large one. The transformer itself is frozen in the final stage.
Add it up on the 1.3B model and 271.8 milliseconds a frame becomes 35.8. On one B200 that is about 28 requests a second, or a 5-second clip in 2.9 seconds. The 14B needs eight B200s to hold 45 milliseconds and 22 requests a second; on a single GPU it falls to 158 milliseconds, which is slower than playback. At 720p the small model needs 127 milliseconds a frame and the large one 116 on eight GPUs, and the paper puts the H100 penalty at 25 to 50 percent on top.
Autoregressive generation is the reason 81 frames is not as limiting as it sounds and also the reason the model cannot do what Ray 3.2 does. Because each block is conditioned on the previous one, clips chain — but chaining is not keyframe control, and there is no path to feed a start frame, an end frame, a reference image or an existing clip. The model has no image conditioning of any kind.


The measurement gap is real, and it runs in both directions
Neither model can be ranked, and the reasons are different enough to be worth separating.
Ray 3.2 has no entry on the Artificial Analysis video arena — not on text-to-video, not on image-to-video, checked 18 September 2026 for this blog. What Luma publishes instead is an internal Ray3 evaluation report claiming parity with Veo 3 and a lead over several rivals, with no numeric scores attached, and a claim that its API runs at less than half the price and latency of comparable models, also unquantified. Those are positioning documents. They are not comparable to an arena Elo with a sample count and a confidence interval, and no page should present them as if they were.
FastFrames has the opposite problem: it has numbers, and all of them come from its own authors. The VBench totals in Table 1 of the paper are 84.23 for FastFrames-S and 84.44 for FastFrames-L, against 84.11 for Causal Forcing++, the baseline it was distilled from, at 272 milliseconds a frame. Wan2.1 1.3B, the base model, scores 83.80 at 30,444 milliseconds. Wan2.2 14.3B scores 84.65 at 117,734. Nobody independent has reproduced any of it, and VBench is a benchmark that rewards exactly the temporal smoothness autoregressive models are good at. The paper's own table is a claim about parity across a 83-to-85.5 plateau, not a measurement of quality.
What the two gaps have in common is the thing to take away: the spread in VBench totals across every model on Apple's list is about two and a half points, while the latency spread across the same list is four orders of magnitude. Whatever that benchmark measures, it stopped discriminating between these systems some time ago. Both of these models are being sold on axes VBench does not see.
Costing one shot in each
Put a concrete brief against both rate cards and the arithmetic stops being comparable at all.
• A five-second 1080p HDR clip — Ray 3.2, roughly $1.20 at listing rates before the HDR multiplier, delivered as EXR that goes straight into a grade. FastFrames cannot produce this shot at any price.
• A five-second 480p silent clip — FastFrames, one B200 for 2.9 seconds of wall clock on the small model, or eight B200s for 3.6 seconds on the large one. Ray 3.2 will not generate at 480p at all; its floor is 540p.
• Ten seconds of anything — Ray 3.2 only from text, at 1080p or below. FastFrames reaches ten seconds by chaining two 81-frame blocks and hoping the seam holds.
• Ninety takes of the same shot — this is where the economics invert. Ray 3.2 charges per clip, so a hundred attempts cost a hundred clips. FastFrames charges GPU-seconds, so a hundred attempts cost a hundred times 2.9 seconds of a machine you already rent.
• The hardware you do not have — FastFrames requires Linux x86-64, an NVIDIA GPU and CUDA 12.8, and the repository ships a working inference path, a latency benchmark script, a VBench evaluation workflow and a demo app. None of that is a hosted service.
That ninety-takes line is the real distinction. Per-clip billing prices a deliverable; per-GPU-second billing prices exploration. If your process is "generate until it is right", the second model is cheaper by an order of magnitude and worse on every other axis. If your process is "generate one thing that has to survive a grade", the first model is the only one of the two that can do the job.
The licence decides more than the spec sheet
FastFrames' weights carry Apple's AMLR terms, and the LingBot-World v2 demo components in the same repository are CC BY-NC-SA 4.0 — non-commercial. The rest is under Apple's Sample Code Licence. "Open weights" and "free to commercialise" are not synonyms, and this release is deliberately careful about the difference in a way that Ray 3.2, sold under a commercial API contract with a Build tier and a Scale tier with an SLA, simply is not.
So the procurement question runs the opposite way to the technical one. Ray 3.2 asks you to accept a vendor relationship: rate card, terms of service, throughput commitment if you buy Scale, an endpoint that may change under you. FastFrames asks you to read three licences and take responsibility for the GPUs. For a studio that already has a legal review process for API vendors, the first is easier. For a research group that already has B200s and a compliance officer who reads licences for fun, the second is.
Where each one belongs, and the question that would change this
Put Ray 3.2 at the end of the pipeline and FastFrames at the start of it, and the two stop competing. Ray 3.2 is what you reach for when the edit exists, the colour pipeline is real, and the shot has to come out the other side gradeable at 10 bits or better — accepting 1080p, five-second HDR clips, no generated audio, and a per-clip bill that multiplies with dynamic range. FastFrames is what you reach for when the loop is the product: a live preview, a controllable world model, an avatar that responds inside a session — accepting 480p, silence, no reference inputs, and a licence you have to read before you ship.
Neither of these is a route on OrcaRouter, and this piece does not claim otherwise. What one key does reach is the hosted video layer that already works today — MiniMax H3 at $0.08 per second at 768p and $0.13 at 2K, and the Kling 3.0 family — alongside 200-plus language models, at provider list pricing passed through with 0% markup so a vendor rate cut lands here the same day it is announced, with automatic failover if an endpoint degrades mid-queue. For a team evaluating either of the models above, that is the cheap way to measure what a real latency and cost profile looks like before committing GPU budget or a Scale contract to a hypothesis.
The thing that would settle this comparison is not a better demo from either side. It is one independent measurement of Ray 3.2 on the image-to-video arena, so the finishing model has a number at all, and one reproduction of FastFrames' 84.4 by somebody who does not work at Apple or Cornell. Until both exist, the honest basis for choosing is the deliverable — and on that axis these two models are not close enough to be confused.

