Topaz Video AI Models Compared: Astra 2, Starlight, Proteus, Gaia, Apollo, Hyperion (2026)
Last updated August 7, 2026
Topaz ships specialist video models, and choosing correctly is the whole game: Astra 2 for creative diffusion upscaling of stylized AI video, Starlight Precise 2.5 to de-plastify GenAI faces and fabrics, Proteus Natural for a no-knobs artifact-free 2x, Gaia 2 for frame-independent animation upscaling, Apollo for slow motion, and Hyperion 2.5 for SDR-to-HDR with ProRes/EXR outputs. All run inside the invideo agent's Tools tab.
Updated August 2026
Topaz Labs doesn't ship one video model — it ships a bench of specialists, and picking the wrong one is the most common reason an upscale disappoints. As of August 2026 the working set is: Astra 2 (creative diffusion upscaling for AI video), Starlight Precise 2.5 (realism restoration), Proteus Natural (fixed 2×, no knobs), Gaia 2 (frame-independent animation 2×), Apollo/Apollo Fast (frame interpolation), and Hyperion 2/2.5 (SDR-to-HDR). This guide is the decision matrix: which model for which footage, with specs and honest caveats from Topaz's own developer documentation.
Which Topaz model should you use?
Match the footage, not the marketing. This table condenses Topaz's developer docs (fetched August 2026) into routing logic:
| Your footage | Pick | Why |
|---|---|---|
| AI-generated, stylized, needs detail invented | Astra 2 | Prompt-guided creative diffusion with a creativity slider |
| AI-generated, must look real (faces, skin, fabric, logos) | Starlight Precise 2.5 | Built to reduce the "plastic" look of GenAI video |
| Real footage, conservative clean 2×, zero artifacts | Proteus Natural | Fixed 2×, no parameters, halo- and aliasing-free by design |
| Real footage needing >2×, tunable sharpening, up to 8K | Proteus (v4) via the Topaz Video Upscaler | Up to 8× with the full slider set |
| Animation, stop-motion, motion graphics, screen recordings | Gaia 2 | Frame-independent — no temporal smearing across frames |
| Slow motion or fps conversion at exact multiples | Apollo (or Apollo Fast for simple motion) | Motion-compensated interpolation at 2×/4×/8× |
| SDR master → HDR deliverable (PQ/ACES pipeline) | Hyperion 2.5 (2 if your pipeline is pinned) | 12-bit ProRes and 16-bit EXR professional outputs |
| Heavily degraded archival (VHS-era noise) | Starlight family (HQ) or Artemis/Nyx | Diffusion restoration vs. classic denoising |
The pattern in that table: only the first two rows ask what you want; every row below routes on what the footage is. AI-generated clips split by intent — Astra 2 to invent, Starlight to stay faithful — and everything else is a footage-type lookup.
What is Topaz Astra 2 best for?
Astra 2 is Topaz's creative diffusion upscaler — "diffusion-based upscaling for GenAI video," per the official docs, generating new detail and texture rather than merely enlarging pixels. Its signature control is a required creativity slider (0–1) plus an optional text prompt steering what gets invented; sharp and realism sliders round out the set. Per the documentation, it performs best on stylized content, wide shots, and detail-sparse scenes.
The honest note, straight from Topaz: Astra 2 is explicitly not for photoreal fidelity work. If the output must stay faithful — a client's face, a product label — creative diffusion is the wrong tool, because it invents. Specs as of August 2026: jobs cap at 9,000 frames without a prompt, 450 with one; cost is 10 frames per credit at 1080p, 6 at 4K. In Topaz's Astra web app, the Precise render mode hands fidelity jobs to Starlight instead.
What does Starlight Precise 2.5 do?
Starlight Precise 2.5 (slp-2.5) is the fidelity counterpart: a diffusion enhancement model "designed to improve realism in AI-generated video by reducing plastic or artificial artifacts," per Topaz's docs. In practice it strengthens exactly what AI generators get subtly wrong — faces, fabrics, material response, small text and logos. Topaz also recommends it for modern archival footage (2000s-era digital video) with enough information to rebuild from.
Specs as of August 2026: default 1080p output with adjustable dimensions, up to 9,000 frames per job, 8–16-bit depth, FFV1/ProRes/VP9 output. Cost runs roughly 26 frames per credit at 1080p (12 at 4K) — cheaper per frame than Astra 2, which surprises people. The wider Starlight family (HQ, Mini, Sharp, Fast 2) covers heavier restoration; 4K export shipped in March 2026.
When is Proteus Natural the right call?
When you want nothing to go wrong. Proteus Natural (pnat-1) is a fixed 2.0× upscaler with zero tunable parameters — deliberately. Topaz's docs say it "eliminates over sharpening and over enhancement," reducing halos, aliasing, and facial distortion, especially on small or background faces, with output slightly softer and more natural than classic Proteus.
That makes it the safe default in multi-step pipelines: a model that can't be mis-tuned is a feature, not a limitation. Pricing is flat: 1 second costs 2 credits at any resolution, per Topaz's published rates. Need more than 2× or manual sharpening control? Step up to Proteus v4 via the Topaz Video Upscaler instead.
Why does animation need Gaia 2?
Because temporal blending — the thing that makes live-action upscalers smooth — actively damages frame-based art. Gaia 2 (ganim-1) is, per Topaz, a "2× upscaling model designed specifically for animation and frame-based sources": it processes each frame independently, with no temporal blending, so stop-motion pops, cel-animation line boil, and motion-graphics hard cuts survive intact instead of smearing into their neighbors.
It's fixed 2× with no parameters, and dramatically cheaper than the diffusion models: 300 frames per credit at 1080p (100 at 4K) — roughly 10–30× less credit burn per frame than Astra 2. For anime, claymation, or screen recordings, Gaia 2 is the better result and the better invoice.
What do Apollo and Apollo Fast do?
Apollo (apo-8) is Topaz's motion-compensated frame interpolation model for slow motion and fps conversion. The docs credit it with handling "non-linear or slightly blurry motion reliably," and recommend retiming at exact multiples — 2×/4×/8× — via slowmo (1–16), target fps (15–240), and duplicate-frame parameters. Apollo Fast (apf-2) covers simpler motion at roughly half the credit cost.
Concrete costs from Topaz's docs: a 10-second 30fps clip at 2× slow motion runs about 1 credit at 1080p, 5 at 4K; at 8×, 9 and 35. For arbitrary retimes, Topaz points to Chronos.
What is Hyperion, and who needs SDR-to-HDR?
Hyperion 2 (hyp-2) automatically converts SDR video to HDR — redistributing luminance and color while improving text, small faces, and fast motion — with genuinely professional outputs: 12-bit ProRes in BT.2020 PQ, or 16-bit float EXR in ACES2065-1. Hyperion 2.5 (hyp-2.5), the current version, adds richer color handling and a wider nit range, and notably converts 8-bit GenAI video to 10-bit ProRes or 16-bit EXR deliverables.
Those pro formats are the point, and also the boundary. For a web MP4 deliverable, HDR conversion buys you little; in an ACES pipeline or PQ mastering, Hyperion is the only model here that speaks your format natively. Both versions cost about 13 frames per credit at 1080p and 6 at 4K.
What about the general Topaz Video Upscaler?
The catch-all Topaz Video Upscaler entry runs Proteus v4 for upscaling with optional Apollo v8 interpolation in one pass — up to 8× and 120 fps output — and exposes the deep catalog (Artemis, Gaia, Nyx, the Starlight line) plus manual sliders for noise, halo, compression recovery, and detail. It's the pick when you want the classic Topaz control surface rather than a specialist's opinion.
What does Topaz cost through the official API?
Topaz publishes API pricing directly (topazlabs.com/api, August 2026): a free Starter tier at $0.12/credit, Developer at $50/month with 500 credits included ($0.10 per extra), Scale at $240/month with 3,000 credits ($0.08 per extra), and custom Enterprise with on-prem options. The Astra web app runs $19–$299/month by credit volume; the desktop Topaz Video app starts at $59/month with unlimited local rendering.
What jobs does the suite actually solve?
- Rescuing a soft AI-generated clip before it ships — the core upscale video job, routed to Astra 2 or Starlight depending on whether you want invention or fidelity.
- General cleanup of real footage — the everyday video enhancer workflow, where the Proteus variants live.
- Finishing passes in a VFX stack — slow-motion retimes and HDR masters sit naturally at the end of an AI VFX pipeline, after compositing.
Topaz questions, answered
Which Topaz model is best for AI-generated video?
Two, depending on intent: Astra 2 to invent detail and push style (it takes a text prompt), Starlight Precise 2.5 to make the clip look more real — less plastic skin, cleaner fabric and text — without changing what's there.
What's the difference between Astra 2 and Starlight Precise 2.5?
Astra 2 is creative: prompt-guided diffusion that generates new texture, best on stylized footage, and explicitly not for photoreal fidelity per Topaz's docs. Starlight Precise 2.5 is corrective: it restores realism while staying faithful to the source.
Is there a Topaz model for anime and animation?
Yes — Gaia 2, which upscales each frame independently with no temporal blending, so frame-based art doesn't smear. Fixed 2×, and among the cheapest models in the suite per frame.
Can Topaz do slow motion?
Apollo handles slow motion and fps conversion at 2×/4×/8× multiples up to 240 fps; Apollo Fast covers simpler motion at about half the credit cost. For non-multiple retimes, Topaz recommends Chronos.
What is Topaz Hyperion used for?
Converting SDR video to HDR with professional outputs — 12-bit ProRes BT.2020 PQ or 16-bit EXR ACES2065-1. Hyperion 2.5 also upgrades 8-bit AI-generated video into 10-bit/16-bit deliverables.
How much does the Topaz API cost?
Official tiers as of August 2026: free Starter at $0.12/credit, Developer at $50/month (500 credits), Scale at $240/month (3,000 credits), Enterprise custom. Per-model burn varies widely — Gaia 2 is roughly 30× cheaper per 1080p frame than Astra 2.
Where can you run the Topaz suite?
The full lineup above — Astra 2, Starlight Precise 2.5, Proteus Natural, Gaia 2, Apollo and Apollo Fast, Hyperion, and the general Topaz Video Upscaler — sits in the Tools tab of the invideo agent, alongside the generation models that produce the footage in the first place. That adjacency is the practical win: generate a shot with Veo or Kling, then hand it to Starlight or Astra 2 without leaving the project. On the AI video platform that gives serious creatives every major model in one place, the enhancement pass is a sentence, not an export.
Version history: first published August 2026, per Topaz developer docs fetched August 2026 (Starlight 4K export: March 2026).