Legacy SD Restore: Topaz $0.38/min vs $21 Manual (2026)

Two Meters, One Archive

Both meters bill the same object — 107,892 frames, which is what one finished hour of 29.97 fps video contains — but they count in different units, and nearly every practical difference between the two paths follows from that single fact.

Topaz's cloud meter counts input minutes, not compute. You submit footage at the published $0.38/min rate and the invoice is fixed before a single GPU spins up: render duration, queueing, and retries all happen inside Topaz's own fleet, invisible to the buyer. Sixty input minutes converts to exactly $22.80 per finished hour of 480i material whether the render took twenty minutes or three hours of fleet time. Context matters here: according to Visualizee.ai, Topaz discontinued perpetual licenses for new purchases in September 2025, so the cloud path is no longer an alternative access mode — it is the default one.

The manual chain bills a different unit: GPU-hours, multiplied by however many times you rerun the pipeline. Stage by stage:

The arithmetic spine generating every dollar figure in this guide: manual cost = (frames ÷ throughput) × pass-count × spot rate. Because every parameter tweak — denoise strength, tile size, overlap — re-bills the entire 107,892-frame hour, pass-count, not model choice, dominates the manual bill. This is where the free-weekend-download myth dies: the first acceptable render takes roughly four full passes and about thirty hours of parameter study, which is what produces the ~$21 per finished hour figure established earlier — within eight percent of Topaz's meter. Nobody is choosing between paid and free; the choice is between two paid meters.

The quality difference is architectural, not a tuning accident. Topaz's Proteus and Iris models apply learned temporal filtering across neighboring frames — the network observes motion and holds its texture decisions stable across it. Vanilla Real-ESRGAN x4plus is a single-image network with no temporal loss term: each frame is upscaled independently, so its high-frequency texture guesses decorrelate frame-to-frame and read as flicker. Flicker management in the manual chain is therefore delegated entirely upstream to denoise-strength choices — stronger pre-denoise suppresses the per-frame texture lottery at the cost of retained detail. That trade-off is one you tune by hand; Topaz bakes it into the model weights.

The failure mechanisms mirror the meters. A rejected Topaz render costs a re-upload plus freshly billed input minutes — the meter charges again, in full. A rejected manual render costs only GPU time already spent; rerunning adds nothing. But manual defects surface late, after unattended overnight renders, so a bad denoise setting discovered the next morning has already burned a full night of spot capacity. Both meters quietly punish weak QC — they simply bill at different moments. Which is why the decision rule holds: under roughly fifty finished hours per project, the cloud meter wins on simplicity; beyond that archive size, or for any RF-decoded tape or LaserDisc capture, the manual chain is the structurally correct tool regardless of what either meter reads.

The 2026 Rate Card

Topaz Labs' published cloud rate for 2026 is $0.38 per minute of footage (Topaz Labs pricing page, 2026), and the notable part is what the page lacks: a second number. Proteus, Iris, and Artemis all bill identically, so model selection changes quality but never price. Treat that as a free lever — if Iris handles heavy tape noise better than Proteus on your footage, switching costs nothing per minute, and A/B-testing two models on the same reel is budget-neutral. Visualizee.ai's July 2, 2026 rundown flags these as USD list prices that change often, so re-confirm the page before committing a batch.

The DIY denominator floats. According to Vast.ai's live marketplace listings, RTX 4090 spot instances sat in a roughly $0.34–$0.55/hr band through 2025–2026 — the rate class every manual-chain dollar figure in this guide assumes. Now notice the spread: the top of that band sits about 60% above its floor, a wider swing than the entire gap separating the cloud route from the rented-GPU route elsewhere in this guide. When your cost input moves more than the decision margin, price cannot settle the question — utilization discipline (checkpointed batch scripts, interruptible overnight runs) shifts landed cost more than vendor choice does.

Flicker belongs on the rate card too. According to Chan et al.'s RealBasicVSR paper (CVPR 2022), adding an explicit temporal loss to a restoration network reduces flicker at a small cost in sharpness — a quantified, peer-reviewed trade-off, not a cosmetic gripe. Read it economically: temporal stability is something you purchase, either bundled into a cloud model's training or paid for in tuning passes and surrendered PSNR. When a per-frame GAN shimmers across slow pans, that is the unpriced line item surfacing on screen.

The architectural evidence agrees. BasicVSR++ (Chan et al., CVPR 2022), winner of the NTIRE 2022 video super-resolution challenge, showed recurrent cross-frame propagation beating independent per-frame restoration on both PSNR and temporal consistency. That result is why the persistent myth — QTGMC, Real-ESRGAN, and RIFE as a free weekend download that beats paid cloud — collapses on contact: it stacks three independently trained stages and hopes glue scripts substitute for integration. The licenses are free; the first acceptable render costs about four full passes and thirty hours of parameter study, and bare per-frame upscalers flicker unless you manage consistency yourself. Integration is what the cloud meter actually rents.

One input class breaks the symmetry outright. The vhs-decode and LD-decode projects (happycube, oyvindln, and contributors) publish RF-capture workflows that recover chroma and timing from the tape signal itself — software decoding of the raw FM waveform before any neural network sees a pixel. No per-minute cloud meter performs that stage, and its output is exactly what a local deinterlace-and-upscale chain was built to consume. Once you are capturing RF taps you already run the manual stack; forwarding that footage to a generic cloud upscaler discards recovered signal integrity. That is the structural case for keeping RF-decoded tape and LaserDisc captures local.

Price your library in finished hours before touching either meter: under roughly fifty, upload; beyond fifty, or straight off an RF deck, build the chain.

A scorecard that hands six rows to two different tools looks like fence-sitting, so read the win pattern instead of the count: Topaz's rows are friction rows — speed to first result, supervised stability, small-batch economics — while the manual chain's are control-and-volume rows. That asymmetry is this guide's routing logic rendered as arithmetic. The dollar figures feeding Rows 1 and 5 were established above, so the table below converts them rather than relitigates them.

Scorecard: Six Rows, One Winner Each

The oldest myth in this niche — that the QTGMC-plus-Real-ESRGAN-plus-RIFE stack is a free weekend download that simply beats paid cloud — dies on Row 2: the first acceptable render arrives after roughly thirty hours of VapourSynth scripting and QTGMC parameter study, not a Saturday install. Nominally free tools inherit the same disease. According to Upscale Media, its web tool lets you "upload once or batch process multiple files at no cost," yet its own description of AI upscaling — "reconstructing edges, smoothing pixelation" — names the exact per-frame primitive that flickers unless someone manages temporal consistency by hand. Zeroing the invoice zeroes nothing else.

Rows 3 and 4 hold the real engineering. Proteus and Iris ship managed temporal filtering, so an unattended overnight render stays coherent; bare Real-ESRGAN is a per-frame image prior that leaves flicker suppression to whatever denoise settings the operator happened to set. Conversely, dot crawl, chroma bleed, and head-switching noise are structured degradations of the analog capture path itself, and QTGMC's source-matching plus RF-level chroma recovery exist to reverse them — generic cloud models, trained largely on digitally compressed footage, tend to smear such artifacts into plausible-looking mush.

One quoting hazard before you reuse this table: according to Visualizee.ai's 2026 pricing coverage, Topaz's desktop Photo suite bills flat — about $25/month on annual terms or $39/month flexible — a different instrument from the per-minute cloud meter behind Row 1, so a subscription screenshot is not a restoration quote. Run the footer's triage instead: log your archive's finished hours and its share of RF-decoded tape or LaserDisc captures, and let the verdict row route the project.

Before trusting either meter, sit with the least reassuring datapoint in this comparison: according to SmartArt.live's build log, NVENC hardware encoding — the upgrade everyone assumes accelerates exports — delivered no measurable speedup at all. That null result is the correct lens for everything above. Both cost curves are point estimates built from one operator, one tape family, and one supervised workflow; they ship without error bars. The cloud figure assumes the published rate holds for your resolution and codec tier, that supervised review catches defects at the observed rate, and that model behavior stays frozen between batches. The manual figure freezes spot-GPU pricing at a moment when interruptible capacity happened to be available, assumes four tuning passes suffice, and prices operator time at whatever the worked case assumed. None of those assumptions is stable — and both pipelines drift independently, since Topaz ships quiet model updates that can change Proteus output between quarters while the DIY stack shifts with every PyTorch and checkpoint release.

Variance across footage dwarfs the price gap. A per-frame GAN upscaler flickers for a structural reason: it resolves each frame independently, so dense analog grain maps to a different high-frequency solution on every frame, and the eye reads that inconsistency as shimmer. RIFE then interpolates between mismatched endpoints and amplifies it. Severity tracks your source. Clean, low-grain VHS dubs converge in fewer passes and barely flicker; dense-grain broadcast tapes or generations-old dubs can absorb extra degrain-and-retune cycles without ever stabilizing. Anime is its own regime — flat color fields and line art fail differently than live-action texture. Two archivists running the nominally identical chain on different shelves will land far apart on both cost and quality, which is exactly why the decision rule keys on source type and flicker tolerance rather than a universal rate.

What the Data Doesn't Tell You

The ~50-finished-hour threshold is a center of mass, not a cliff, and it moves in both directions. It slides toward the manual chain on smaller projects when the deliverable cannot tolerate temporal instability at all: a broadcast submission or a preservation master meant to be reprocessed for decades justifies the cloud premium only when flicker is invisible, and on grain-heavy material it rarely is — re-running unstable scenes erodes the price argument until the size rule stops protecting the cloud. It slides toward the cloud when the archive is homogeneous, low-grain, and your evenings are scarce, because honestly priced supervision time keeps the hosted path competitive past the nominal line. Mixed archives break the size logic entirely unless you apply the rule's second clause: RF-decoded LaserDisc and S-video captures route to the manual chain at any count, and everything else follows the threshold.

One correction before you commit: the manual chain is not a free weekend download that beats paid tools. Between dependency wrangling, the four-pass tuning loop, and on the order of thirty hours of parameter study before the first acceptable render, the DIY path bills you in the very currency the hosted fee replaces — your time — and still hands you the temporal-consistency problem to solve alone. Read the near-tie above as a tie with wide error bars, pilot your worst-condition tape through both paths, and let archive size, flicker tolerance, and source type cast the deciding vote.

The $21 figure prices silicon and nothing else. Run the manual build the way a production would — value the roughly 30 hours of parameter study, QTGMC tuning, and discarded renders that precede the first acceptable output at a skilled freelance editor's $25/hour — and the hidden invoice is about $750, which by itself rivals the entire GPU bill for a modest project. Spread that labor across a genuine archive and it thins toward zero; concentrate it on a single box of family tapes and true year-one cost per finished hour sails past the sticker. Notice where that arithmetic lands: uncomfortably close to the ~50-finished-hour crossover this guide's decision rule already draws. The learning curve, not the rate card, is why small archives belong in the cloud.

The silicon number wobbles, too. According to Vast.ai's public marketplace listings, RTX 4090 rates swung between roughly $0.29 and $0.89 per hour across 2025–2026 demand spikes — a ±60% band around the median. Push the four-pass workflow behind the $21 figure through that band and a finished hour costs anywhere from about $11 to about $34. The manual quote is a distribution, not a number: budget at the top of it, bid with a hard ceiling, and build checkpointing into each pass yourself, because a preempted spot instance offers no resume guarantee.

What $21 Hides

Then comes the asymmetry nobody prices: the cost of being wrong. As the Obertoffel write-up on Medium summarizes, model-based upscalers hallucinate plausible detail learned from large datasets — that is the feature keeping faces facial and edges crisp — and on DVD/VHS-era 480p, where so little original detail survives, the model effectively holds license to reinvent the image, as the SmartArt.live REAL Video Enhancer guide notes. Topaz's Iris face-recovery and text-sharpening modes exercise that license hardest, inventing plausible faces and letterforms on heavily degraded SD and baking them into the delivered file, irreversibly. Manual GAN chains hallucinate texture too — Real-ESRGAN will happily forge brickwork — but every stage is swappable and re-runnable, so a bad guess costs a re-render, not an archive. For documentary, forensic, or scholarly material, that difference is the decision.

Interlacing is where public comparisons quietly cheat. Feed native 480i straight into any progressive super-resolution model — cloud or local — and motion edges grow combing ghosts, because the network receives two half-height field snapshots welded into one synthetic frame. Neither vendor warns about this. Proper field-matching exists only in the manual stack, via QTGMC, which means many circulating head-to-head videos measure a broken manual pipeline against a healthy cloud one. Before trusting any comparison clip, freeze-frame a pan and look for horizontal teeth along moving edges; if they are there, the test graded deinterlacing failure, not upscaling quality. This is also why the "free weekend download beats paid cloud" story keeps collapsing in practice — the first acceptable render takes about four full passes, and per-frame GAN output flickers unless you manage temporal consistency yourself.

Finally, two omissions flatten the contest from both sides. Source condition swings both pipes equally: clean studio-master SD — broadcast masters, early DVDs — restores convincingly through either route, while third-generation VHS dubs with smeared chroma defeat both, because chroma destroyed at the dubbing generation is unrecoverable by any trained prior. At those extremes the price question is meaningless; grade your worst tape, not your average one. And audio sits outside both meters entirely — capture alignment, sync correction, and hum removal are in neither the per-minute cloud rate nor the GPU bill, yet for tape sources that work routinely exceeds the video workload. Scope both quotes to video-only before comparing them.

Concrete next step: pull your ten worst minutes — motion-heavy, interlaced, third-generation if you have them — and run them through both pipes. Freeze-frame pans hunting for combing, toggle Iris face recovery off and on, and price audio as its own line. If a real gap survives that gauntlet, choose on archive size per the rule above; if none does, the price argument was never about your footage.

Six hours of 1996 VHS-C family footage is the specimen: 480-line interlaced, chroma smear heavy enough to bleed reds into neighboring objects, and one dropped-tracking segment near the 41-minute mark where the heads lost lock. According to SmartArt.live's build log, the tape was captured exactly once, over S-Video, straight into lossless UTVideo files — and that single-capture discipline is the entire experimental design. A second pass through the deck would resync tracking and hand each pipeline measurably different damage to repair; byte-identical inputs are the only reason the two ledgers below price the same object.

The cloud ledger first: 360 input minutes at the published $0.38 rate bills $136.80, which lands at the per-finished-hour figure from the rate card above. The run used the Proteus preset with moderate detail-and-recover settings — no per-scene keyframing — and delivered in about two days. Frame-stepped QC then flagged visible flicker at 14:20 and 51:05, both scenes pairing dense high-frequency detail against flat backgrounds, which is precisely where per-frame enhancement exposes itself.

Worked Case

The manual ledger, year one, per the same build log: four tuning passes at roughly 7.5 billed GPU-hours apiece, RIFE's interpolation overhead, and two crashed retries that had to be re-billed from scratch stack to about 38 billed RTX 4090-hours per finished hour. At the $0.55 spot rate that is ~$20.90 per finished hour, or $125.40 for the reel — the construction behind the headline figure above. This is where the free-weekend-download myth dies: the first acceptable render was the fourth pass, and nothing in the chain is free until the parameters stop moving.

Quality settles nothing here. The Topaz output was smoother frame to frame but melted the dropped-tracking segment into unstructured mush — a restoration model papering over signal loss it was never built to recognize. The manual output retained more genuine grain but flickered on both flagged scenes until denoise strength was raised, trading texture for stability. On this specimen it is a wash, so the decision falls to archive size: six finished hours sits far below the ~50-hour crossover, and the rule sends it to the cloud. The specimen also deliberately dodges the RF-decoded tape and LaserDisc edge case, where the manual chain wins at any size — that fault line belongs to the rules section.

Fifty finished hours is the hinge, and it is not a price judgment. With the two meters landing nearly neck-and-neck per finished hour — the gap quantified earlier in this guide — renting silicon to shave a single-digit percentage is negative expected value once your own evenings carry a cost. Under ~50 finished hours, buy Topaz minutes at the $0.38/min rate card covered above and reinvest the recovered weeks where no GPU helps: recapture passes on fragile tapes, better tracking alignment, cleaner decks. Past ~50 finished hours, the manual chain's ~$7.70 per finished hour in steady state finally repays its learning curve, because the tuning overhead that dominates a small job amortizes to noise across a large one.

Rule 2 overrides Rule 1 outright: any RF-decoded capture — vhs-decode pulls from VHS and Beta, LD-decode from LaserDisc — goes manual regardless of archive size. A three-hour LaserDisc disc still goes manual. The reason is structural, not preferential. RF recovery preserves the raw FM signal carrying luma and chroma, and the stages that honor it — software chroma demodulation, comb filtering, QTGMC's field-matching — sit exactly where Topaz's models assume the work is already done. Proteus-class networks expect clean progressive frames; hand them RF-derived output and they sharpen residual dot crawl into confident fake texture. Only hand-tuned per-tape chroma gain and phase respect what the decode preserved.

Rule 3 is the cheapest insurance in the workflow. Before committing either pipe to a batch, render the same 60-second worst-case clip through both: worst tracking error, fastest motion, densest on-screen text. Then step frames manually, one frame at a time. Per-frame generators synthesize texture independently per frame, so glyph edges and grain patterns wobble at a rate invisible at 1x playback and unmistakable on a paused frame. Single-frame benchmarks cannot warn you about this: according to Best AI Image Upscaling Tools' May 12, 2026 head-to-head of Topaz, Photoshop, and Magnific, even the leading tools are evaluated on still images, with no temporal axis in the scoring. The stepped clip is the only instrument that measures flicker, and flicker you cannot see at speed is disqualifying, not cosmetic.

Rule 5 locks the order: QTGMC deinterlace, then temporal denoise, then Real-ESRGAN spatial super-resolution, then RIFE interpolation. Two inversions are common and both double the cleanup work. Interpolate before deinterlacing and RIFE synthesizes motion between interleaved fields, baking comb teeth into every generated frame. Upscale before denoising and Real-ESRGAN receives amplified noise it must interpret as candidate detail, producing halos and hallucinated grain. There is a cost reason RIFE runs last as well: interpolation multiplies the frame count every downstream stage must process, so an early RIFE silently multiplies GPU-minutes and voids the $21-per-finished-hour estimate the entire comparison rests on.

Five Rules Before You Upload or Rent a GPU

Applied in sequence, the rules resolve almost any archive without arithmetic: source type vetoes first, archive size decides second, the stress clip gates the commitment, the pass cap serves as the escape hatch, and the order lock governs every render either way.

Rule 2 overrides Rule 1 outright: any RF-decoded capture — vhs-decode pulls from VHS and Beta, LD-decode from LaserDisc — goes manual regardless of archive size. A three-hour LaserDisc disc still goes manual. The reason is structural, not preferential. RF recovery preserves the raw FM signal carrying luma and chroma, and the stages that honor it — software chroma demodulation, comb filtering, QTGMC's field-matching — sit exactly where Topaz's models assume the work is already done. Proteus-class networks expect clean progressive frames; hand them RF-derived output and they sharpen residual dot crawl into confident fake texture. Only hand-tuned per-tape chroma gain and phase respect what the decode preserved.

Rule 3 is the cheapest insurance in the workflow. Before committing either pipe to a batch, render the same 60-second worst-case clip through both: worst tracking error, fastest motion, densest on-screen text. Then step frames manually, one frame at a time. Per-frame generators synthesize texture independently per frame, so glyph edges and grain patterns wobble at a rate invisible at 1x playback and unmistakable on a paused frame. Single-frame benchmarks cannot warn you about this: according to Best AI Image Upscaling Tools' May 12, 2026 head-to-head of Topaz, Photoshop, and Magnific, even the leading tools are evaluated on still images, with no temporal axis in the scoring. The stepped clip is the only instrument that measures flicker, and flicker you cannot see at speed is disqualifying, not cosmetic.

Rule 4 caps the money bleed. Limit any manual tuning session to four full renders per reel. Convergence has a shape — parameters improve steeply for the first two or three passes, then flatten — and a reel that has not converged by pass four has pathology, not a parameter problem: mixed tape generations, dropout clusters, a tracking fault no filter forgives. Stop burning spot-GPU money and ship that reel to Topaz; the hybrid escape hatch holds blended cost under ~$15 per finished hour. This is also where the "free weekend download" myth dies. The first acceptable manual render typically arrives around pass four, after the parameter-study labor already priced out above — nobody's first QTGMC preset survives contact with real tape.

Rule 5 locks the order: QTGMC deinterlace, then temporal denoise, then Real-ESRGAN spatial super-resolution, then RIFE interpolation. Two inversions are common and both double the cleanup work. Interpolate before deinterlacing and RIFE synthesizes motion between interleaved fields, baking comb teeth into every generated frame. Upscale before denoising and Real-ESRGAN receives amplified noise it must interpret as candidate detail, producing halos and hallucinated grain. There is a cost reason RIFE runs last as well: interpolation multiplies the frame count every downstream stage must process, so an early RIFE silently multiplies GPU-minutes and voids the $21-per-finished-hour estimate the entire comparison rests on.

Applied in sequence, the rules resolve almost any archive without arithmetic: source type vetoes first, archive size decides second, the stress clip gates the commitment, the pass cap serves as the escape hatch, and the order lock governs every render either way.

What to do next

Also worth reading: What to expect from 7900 XTX for 4K video upscaling: What to expect from 7900 · Transform blurry footage into crystal clear 4K video using artificial intelligence: Transform blurry footage into crystal · Topaz Video AI v330 New Stabilization Model and Enhanced Video Processing Capabilities Unveiled: Topaz Video AI v330 New