Flux 3 is Black Forest Labs' third generation model and the first in the line to render video and audio next to images. This review covers what Self-Flow changes, what the published material does and does not say, how it stands beside Runway and Sora, and what access costs.
Flux 3 is the third generation of Black Forest Labs' model family and the first that renders video and audio rather than stills alone. The line has moved quickly: FLUX.1 in 2024 was a 12B rectified-flow image model with an open dev checkpoint, FLUX.2 in late 2025 merged generation and editing behind a larger vision-language encoder, and Flux 3 adds a training scheme called Self-Flow plus two output types new to the family.
This review answers the four questions people bring to a new model: what changed in the architecture, what it can and cannot generate, how it stands next to Runway and Sora, and what access costs. Vendor numbers are labelled as claims, and typical industry ranges are written as ranges.
FLUX.1 arrived in 2024 as a 12B rectified-flow image transformer with a CLIP encoder and T5-XXL, in three tiers: an API-only pro model, a non-commercial dev checkpoint and an Apache 2.0 schnell build distilled to a few steps. FLUX.2, released in late 2025, scaled the family to a 32B model with a 24B vision-language encoder, merged generation and editing into one checkpoint and pushed output towards 4 MP. Flux 3 keeps that direction and pulls video and audio into the same line.
The lab behind it is Black Forest Labs, founded in 2024 by researchers who wrote the latent diffusion work behind Stable Diffusion. That background explains a licensing split that catches people out: hosted models are commercial products, while downloadable checkpoints usually carry a non-commercial licence with a paid commercial agreement beside it. Read the licence file, not the launch post, before the output goes into client work, and check whether the commercial tier covers video as well as images.
Both earlier generations used rectified flow, where training prescribes a mostly straight path from noise to data and guidance is applied at sampling time or distilled into a second checkpoint. Read literally, Self-Flow describes a model that sets that path for itself instead of following a fixed schedule, and the release material calls it a change to training, not to the network itself. The public write-up stops there: objective, schedule and data mix are unpublished, so nobody outside the company can verify the mechanism. Treat the architecture story as a vendor claim and judge the model on reproducible output.
The advertised consequence is operational. One checkpoint is meant to serve both a short draft pass and a longer high-fidelity pass, so you stop juggling a base weight and a distilled twin, and one shared representation covers stills, video frames and audio. That is testable on your own prompts: render at the step counts you actually use and compare how far the composition drifts, instead of reading a step-count chart in a launch post.
Image work is the mature part of the family, steady at prompt adherence and legible text rendering since FLUX.1; later versions added multi-reference editing, where you hand over several images and describe the change instead of painting a mask. Video behaves like other clip models in this generation: five to ten seconds of coherent motion at 720p to 1080p is a realistic target; anything longer comes from stitching rather than one long render. Audio in this class means synchronised sound generated with the frames, which suits a storyboard and does not replace a dialogue edit.
| Output type | Typical result | Where it fits |
|---|---|---|
| Image | Stills and instruction-based edits at a few megapixels | Concept boards, texture sources, 3D reference art |
| Video | Five to ten seconds at 720p to 1080p, 24 or 30 fps | Animatics, motion studies, short social cuts |
| Audio | Synchronised sound rendered with the clip | Rough mixes for review, not finished dialogue |
| Editing | Consistent changes across several reference images | Product variants and character consistency |
Resolution, frame rate and clip length decide whether a model fits a pipeline, and all three shift between versions. The shape of the constraints is more stable: video models in this generation tend to run at 24 or 30 frames per second, at resolutions that are multiples of 16 or 32 pixels, and express duration as a frame count rather than seconds. Self-hosting is a memory question first: an open checkpoint of this size typically wants 24 GB of VRAM or more at reduced precision, or less quantised.
Public leaderboards for generative video mostly measure human preference on short, well-lit clips, and those votes reward polish more than control. A model can top a ranking and still be wrong for a project with a locked camera move and fixed character sheet. The criteria worth scoring are narrow: prompt adherence on your subject matter, identity drift across a clip, how much the endpoint exposes in seeds and reference images, what the licence permits commercially, and cost per usable second rather than per generation. The table below is editorial judgement, not a measured benchmark.
| Criterion | Flux 3 | Runway | Sora |
|---|---|---|---|
| Access | Vendor API, aggregators, open dev weights | Hosted product and API, no weights | Vendor apps and API only |
| Control | Seeds, reference images, instruction-based edits | Camera and motion controls for editors | Few knobs, prompt-to-clip by design |
| Audio | Synchronised sound with the frames | Depends on the version | Native audio is a selling point |
| Commercial terms | Per-call pricing, non-commercial weights | Subscription credits per generation | Subscription tiers by length and resolution |
The vendor's own API and playground is the reference implementation and gets new versions first. Aggregators such as fal and Replicate host the same checkpoints with per-call billing and fast cold starts, handy for prototyping, at the cost of another dependency and more terms. Self-hosting an open dev checkpoint removes the per-call cost and keeps material private, at the price of hardware, setup time and licence limits. Image endpoints in this class commonly land around one to five US cents per image near a megapixel, while video endpoints bill per second of output; treat both as snapshots, not list prices.
| Route | What you pay for | What to watch |
|---|---|---|
| Vendor API | Per image or per second of video | Spend scales with iteration, limits change by version |
| Aggregator | Per call, sometimes per GPU-second | An extra dependency and another set of terms |
| Self-hosted weights | Hardware, power and engineering time | Non-commercial licence on most dev checkpoints |
| Rented GPU | Hourly rental while you work | Cost accrues while debugging, storage is extra |
Version numbers, licences, prices and third-party policies here change every few months. Everything above is a snapshot from the publication date; verify current numbers and terms on the vendor page before committing budget.
For a 3D artist the value is upstream of geometry. Reference sheets, multiview turnarounds, albedo texture sources and fast concept variations are image problems, and a model with strong prompt adherence and multi-reference editing handles them well. It does not produce a mesh: no topology, no UV layout, no millimetre scale and no guarantee of a watertight shell, so the output feeds an image-to-3D step instead of replacing it. Open3DAI works on that step: text-to-3D and image-to-3D generation, multiview input, AI texturing, side-by-side engine comparison, export to GLB, glTF, OBJ, STL, PLY and 3MF, plus 48 format converters, 12 browser tools and a REST API. It does not retopologise, rig, unwrap UVs or promise print-ready meshes without cleanup, and dimensioned parts still belong in a parametric CAD tool.
The family has followed a tiered pattern rather than a single licence: hosted models behind the vendor's API, an open dev checkpoint, and occasionally a permissive build such as the Apache 2.0 schnell release in the first generation. Open dev weights in this line have normally carried a non-commercial licence with a separate commercial agreement, so check the licence file for the specific checkpoint and confirm whether video and audio output is covered.
FLUX.2 was a 32B image model that merged generation and editing and accepted multiple reference images. Flux 3 adds video and audio output to the same line and introduces the Self-Flow training scheme, so the practical difference is not better stills alone but one model covering several output types. Parameter counts, context limits and resolution ceilings move between versions, so read the current model card rather than a summary post.
No. It generates images, video and audio, and the output is pixels rather than geometry. To get a mesh you feed a still or a multiview set into an image-to-3D step and then clean the result in a DCC tool such as Blender. Expect to fix topology, scale and non-manifold edges before printing or importing into a game engine.
Hosted image endpoints in this class commonly run from about one to five US cents per image near a megapixel, and video endpoints bill per second of generated output, with the rate rising at higher resolution. Aggregators add their own margin, while self-hosting trades per-call fees for hardware or hourly GPU rental. Vendors change prices and credit packs often, so treat any figure as a snapshot and confirm it on the pricing page.
Only on the self-hosted route; API and aggregator access runs on the provider's hardware. If you do self-host an open checkpoint of this class, plan on roughly 24 GB of VRAM or more at reduced precision, with quantised community builds fitting into smaller cards at some cost in fine detail. Renting a GPU by the hour is the usual middle path.
Treat five to ten seconds as the realistic band for a coherent shot, at 720p to 1080p and 24 or 30 frames per second. Longer sequences are usually built by extending or stitching shots rather than asking for one long generation, because subject identity and props drift as the clip grows. Some endpoints express duration as a frame count, which is why the arithmetic rarely lands on a round number of seconds.
The fastest route is the vendor's own API or playground, which gets new versions first. Aggregators such as fal and Replicate host the same checkpoints behind one key and are convenient for side-by-side tests. Open dev weights, when a tier exists, are usually published on Hugging Face under a non-commercial licence. Availability differs by region and account tier, so confirm what your account can actually call before designing around it.
There is no single winner, and a review that names one without stating criteria is not useful. Flux 3 is attractive when you want editing consistency, reference-image control and the option to self-host; Runway suits editors who want motion presets and a timeline; Sora is a strong route to a polished short clip with sound. Score prompt adherence, identity drift, licensing and cost per usable second on your own prompts, then re-run the comparison after a version update.
Send the same prompt or image to several engines, compare the meshes side by side, then export GLB, glTF, OBJ, STL, PLY or 3MF. Creating an account grants free credits and needs no card.
MiniMax H3 Max is fal's post-trained version of the open-weight MiniMax H3 video model, tuned for prompt adherence and served on an inference stack co-designed with it. Here is what it changes, how to read the 35x throughput figure, what a second of output costs, and when the stock H3 endpoint is still the right pick.
Sulphur 2 is a community fine-tune of Lightricks' LTX 2.3 that runs on your own GPU with no service-side prompt filter. Here is what the download contains, how much VRAM it really needs, how the ComfyUI workflows run, and where it beats or trails Runway and Kling AI.
3D AI Studio sells one credit balance that covers image-to-3D, text-to-3D, texturing, rigging and a shelf of mesh utilities, and it routes part of that work to Meshy, Tripo and Rodin. This review works through the published credit schedule, the list prices behind it, and the pipelines where a subscription is the wrong purchase.