Luma Photon now names two generations: the cheap 2024 Photon/Photon Flash tier ($0.015/$0.004 per image) and the flagship Uni-1/Uni-1.1 — a decoder-only autoregressive model uniting reasoning and generation, with native 2K, up to 9 reference images, multilingual text, and official pricing of $0.0404 (uni-1.1) to $0.10 (uni-1-max) per image. Weakness: ~31-second latency. Both tiers run inside invideo.
Updated August 2026
Luma Photon is Luma's image-model family, and the name now spans two generations. The original Photon and Photon Flash (December 2024) are the cheap, fast workhorses — $0.015 and $0.004 per 2-megapixel image, per Luma's Photon page. The current flagship is Uni-1 and its API revision Uni-1.1: a decoder-only autoregressive transformer that processes text and image tokens in a single interleaved sequence, so reasoning and generation happen inside one model rather than a prompt rewriter bolted onto a diffusion backbone. Uni-1 outputs native 2K, accepts up to 9 reference images, and edits through the same endpoint it generates with. In invideo's model picker the family appears as Luma Photon (uni-1) and Luma Photon (uni-1-max) — Photon is the family name that stuck; Uni-1 is what Luma calls the current models.
Is Luma Photon the same as Uni-1?
Practically yes, technically no. Photon is the umbrella under which Luma's image models are widely listed, but Luma itself brands the current line Uni-1 — announced in early March 2026 (the date comes from Luma's social posts; the product page is undated) and described in Luma's own words as "a multimodal reasoning model that can generate pixels." The older Photon models still exist as a separate, cheaper tier. The variants:
| Variant | Generation | What it is | Official price (text-to-image) |
|---|---|---|---|
| Photon | Dec 2024 | Original quality model, 2MP/1080p output | $0.015 per image |
| Photon Flash | Dec 2024 | Speed-optimized budget tier | $0.004 per image |
| Luma Photon (uni-1) → uni-1.1 | Mar–May 2026 | Current flagship: unified reasoning + generation, native 2K | $0.0404 per image |
| Luma Photon (uni-1-max) | 2026 | Higher-quality tier of Uni-1 at the same 2K resolution, roughly 2.5x the rate | $0.10 per image |
Prices are Luma's official API rates (lumalabs.ai/photon, lumalabs.ai/api, as of August 2026). When a model picker says "Luma Photon (uni-1)," you are getting the Uni-1 generation, not the 2024 model.
Luma Uni-1 specs and architecture
| Spec | Uni-1 / Uni-1.1 |
|---|---|
| Architecture | Decoder-only autoregressive transformer; text and image tokens in one interleaved sequence |
| Resolution | Native 2K (2048px), standard aspect ratios, PNG/JPEG |
| Modes | Text-to-image, image editing, and reference-guided generation through the same endpoint |
| Reference images | Up to 9, usable for identity, composition, or style |
| Languages | Multilingual prompts and in-image text, including Chinese, Japanese, and Arabic |
| Latency | ~31 seconds per image |
| Weights | Closed; API and first-party apps only |
Everything in this table follows from the first row: one autoregressive stream is what buys the instruction-dense editing, the multilingual in-image text, and the 9-reference control — and the ~31-second latency is the bill for it.
Sources: the Uni-1.1 API announcement (May 5, 2026) and Luma's API page, as of August 2026.
How much does the Luma image API cost?
Official pay-as-you-go rates at 2K output (lumalabs.ai/api, August 2026):
| Task | uni-1.1 | uni-1-max |
|---|---|---|
| Text-to-image | $0.0404 | $0.1000 |
| Image edit | $0.0434 | $0.1030 |
| Generation with 1 reference | $0.0434 | $0.1030 |
| Generation with 8 references | $0.0644 | $0.1240 |
Edits and references carry almost no premium — each added reference costs $0.003 — and uni-1-max is about 2.5x uni-1.1 at the same resolution: a quality tier, not a resolution tier. Third-party resellers quote different numbers; these are the only official figures.
What Uni-1 does well
Reasoning and generation in one pass. Because text and image tokens share one autoregressive sequence, the model that interprets your prompt is the model that paints the pixels. Luma's pitch — "a multimodal reasoning model that can generate pixels" (lumalabs.ai/uni-1) — shows up concretely in spatial reasoning, multi-panel and storyboard output, and instruction-dense edits.
Reference-guided work. Up to 9 reference images for identity, composition, or style is more than most image APIs expose, and it is the backbone of character-consistent stills.
Editing precision. Natural-language editing runs through the same endpoint as generation, and on July 29, 2026 Luma shipped Layers — element-level editing "powered by Uni-1" that adjusts individual elements instead of re-rolling the whole frame.
Text and languages. Documented text rendering plus multilingual coverage including CJK scripts and Arabic — rare among image models.
On benchmarks: Luma claims #1 Human Preference Elo for Overall, Style, Editing, and Reference-based generation (#2 for text-to-image), and a lead on RISEBench for reasoning and spatial logic. These are vendor-reported figures from Luma's own product page — treat them as the company's claims, not independent rankings.
Where Uni-1 falls short
It is slow. ~31 seconds per image (Luma's own figure, August 2026) is an order of magnitude behind turbo-class diffusion models. For high-volume drafting, that latency — not the price — is the real cost.
Closed and vendor-benchmarked. Weights are not available, and until third-party arenas publish placements, the #1-Elo story rests entirely on Luma's numbers.
The max tier's value is unquantified. Luma documents uni-1-max as a higher-quality tier at 2.5x the price but publishes no benchmark separating it from uni-1.1 — test both before committing volume.
How to prompt Luma Uni-1
- Write sentences, not tag soup. A reasoning model follows structured instructions better than keyword lists — spell out subject, layout, style, and any in-image text.
- Use references for what they are typed as. A reference can carry identity, composition, or style; say which job each image is doing.
- Ask for panels explicitly. Multi-panel output is documented — specify the panel count and what happens in each.
- Put exact text in quotes. For posters or mockups, quote the literal string you want rendered, including non-Latin text.
- Edit instead of re-rolling. At $0.003 over the base price, an edit that fixes one element beats regenerating the frame.
What can you make with Luma Photon (Uni-1)?
- Production stills and marketing images. Native 2K output slots straight into an AI image generator workflow without an upscaling pass.
- Storyboards and multi-panel sequences. The documented multi-panel mode plus spatial reasoning makes it a natural engine for storyboard work — and those panels double as keyframes for video.
- Consistent characters. Nine reference slots for identity and style are exactly what serialized character generation needs — the same face and style in every frame.
Luma Photon FAQ
Is Luma Photon the same as Luma Uni-1?
Uni-1 is the current generation of the family widely listed under the Photon name. Photon and Photon Flash (December 2024) remain available as a cheaper tier; Uni-1/Uni-1.1 (2026) is the flagship. A picker entry like "Luma Photon (uni-1)" runs the Uni-1 generation.
How much does Luma Uni-1 cost per image?
Officially $0.0404 per 2K text-to-image generation on uni-1.1 and $0.10 on uni-1-max; edits and single-reference generations run $0.0434 and $0.1030 respectively (lumalabs.ai/api, August 2026).
How many reference images does Uni-1 support?
Up to 9 per generation, covering identity, composition, and style. Pricing scales gently: eight references cost $0.0644 per image on uni-1.1 versus $0.0404 with none.
Why is Uni-1 slower than other image models?
Uni-1 takes roughly 31 seconds per image (Luma's figure, August 2026) — the autoregressive architecture behind its instruction-following also makes it far slower than few-step diffusion models. Use it for finals and edits, not bulk drafts.
Can Uni-1 render text in languages other than English?
Yes — multilingual prompting and in-image text rendering are documented capabilities, including Chinese, Japanese, and Arabic (Uni-1.1 API post, May 2026).
What is Luma Layers?
An element-level editing capability announced July 29, 2026, powered by Uni-1: it exposes individual elements of an image for precise adjustment instead of whole-image re-generation (announcement).
Where can you use Luma Photon and Uni-1?
Both tiers run inside invideo, listed in the picker as Luma Photon (uni-1) and Luma Photon (uni-1-max) among the platform's 200+ models — so the agent can hand Uni-1 the jobs it wins (reference-locked characters, storyboard panels, edits) while faster models absorb the volume work its latency makes painful. Luma's video sibling Ray is in the same roster — Ray 3.2 sits alongside Veo 3.1 and Kling 3.0 — so a Uni-1 storyboard can flow straight into Ray-generated motion; the full lineup is on the invideo AI models index.
Version history: Photon + Photon Flash launch Dec 2024 → Uni-1 announced ~Mar 2026 (date per Luma's social posts) → Uni-1.1 API with generation + natural-language editing May 5, 2026 → Layers element-level editing Jul 29, 2026.