The Image-to-Video Landscape: Architecture, Models, and Real Prices in 2026

🌏 中文版

Feed a model one image and one sentence, get a moving clip back. That is unremarkable in 2026 — but choosing a model is harder than it was two years ago, because visual quality has stopped being a useful discriminator. On the Artificial Analysis Image-to-Video leaderboard (snapshot of 2026-07-27), the top seven models span an Elo range from 1200 down to 1088, with seven entries crammed into that gap. What actually drives the decision is elsewhere: whether you need native audio, whether you can self-host, how fine-grained the control surface is, and what it really costs per second.

This piece takes the I2V pipeline apart, then talks cost using official pricing pages rather than review-site numbers. Verification turned up three errors that circulate widely in secondary coverage; each is flagged in the relevant section.

1. What makes I2V hard is constraints, not fidelity

Image-to-Video Diffusion: From Foundations to Open Frontiers (arXiv:2605.17248), published in May 2026, is the only survey that treats I2V as a subject in its own right. It states the difference from text-to-video (T2V) up front:

Compared with broader video generation settings, this task places stricter demands on content consistency, identity preservation, and motion coherence.

Unpacked: T2V only has to look good and plausible. I2V carries three extra constraints — content consistency (don't invent things absent from the reference image), identity preservation (faces must not drift), and motion coherence (no inter-frame jitter). This is also why I2V is often more useful in practice than T2V: the reference image constrains the output space, so results are more predictable.

The survey compresses the whole field into four design axes: condition encoding, temporal modeling, noise prior design, and spatial-temporal upsampling. Essentially every difference between models falls along one of these.

2. What the pipeline looks like

reference image ──┐
prompt ───────────┼─→ condition encoding ──┐
                  │                        ↓
        3D causal VAE ─→ latent ─→ DiT denoising ─→ latent' ─→ VAE decode ─→ 480p/720p
         (16x/4x compression)                                                    ↓
                                                        separate super-resolution ─→ 1080p

Three decisions carry the design:

1. Diffuse in latent space, not pixel space. The HunyuanVideo 1.5 technical report (arXiv:2511.18870) gives concrete numbers: an 8.3B-parameter DiT paired with a 3D causal VAE compressing 16x spatially and 4x temporally, with 32 latent channels. Without compression the token count explodes and both training and inference become infeasible.

2. U-Net gave way to DiT. Early work inserted 1D temporal layers into a text-to-image 2D U-Net — pseudo-3D in effect — which let it inherit a mature T2I backbone for free. DiT instead cuts video into spacetime patches, treats them as tokens, and feeds them to a Transformer, buying scalability. The survey documents this migration explicitly, and essentially every production model now sits on the DiT side.

3. Cascade rather than one shot. HunyuanVideo 1.5 runs two stages: the DiT produces 480p–720p at 5 to 10 seconds, then a dedicated video super-resolution network takes it to 1080p. This decouples "is the motion right" from "is the image sharp" — regenerating wrong motion is expensive, while adding sharpness is cheap.

3. Controllability lives entirely in the conditioning layer

How usable an I2V model is comes down almost entirely to conditioning design. The survey identifies four types, in increasing order of control:

If your requirement is "product shot with a specified push-in," categories three and four are what you should be shopping for. Leaderboard Elo tells you nothing about that.

4. The 2026 theme is cutting compute, not adding fidelity

The most interesting movement this year is not in quality but in inference cost, and three lines are pushing at once.

Step distillation. NVIDIA open-sourced Cosmos3-Super-Image2Video-4Step on July 20. The official announcement is blunt about it:

They cut sampling from 50 denoising steps to just 4, and remove the need for classifier-free guidance — which otherwise doubles the work per step — for up to 25× faster inference than the base model, with little to no loss in generation quality.

Note that two things stack here: fewer steps, plus removing classifier-free guidance, which otherwise costs two forward passes per step. The more interesting result is the ranking — NVIDIA reports the distilled model as the top open-weight entry on the Artificial Analysis I2V leaderboard, ahead of its own teacher. The 25x figure is NVIDIA's own claim with no third-party reproduction, so discount it accordingly.

Sparse attention. HunyuanVideo 1.5's SSTA (Selective and Sliding Tile Attention) prunes redundant spatiotemporal kv blocks; the report claims a 1.87x end-to-end speedup over FlashAttention-3 on 10-second 720p synthesis. This is a direct answer to DiT attention cost growing quadratically with token count.

The autoregressive branch has not taken off. In principle, chunk-by-chunk AR generation extends to long clips more cheaply and holds long-range coherence better. In practice errors accumulate along the sequence and per-clip inference is slower. It has not displaced DiT in production; it mostly shows up as an extension feature bolted onto DiT models.

5. Current models, and one shutdown to plan around

Artificial Analysis I2V leaderboard, 2026-07-27 snapshot (Elo, blind preference):

Hold the leaderboard loosely. It measures blind preference, which is not the same as "works well for your case," and it moves month to month. The top two are one point apart, which is a tie in any practical sense.

Sora is gone. This is the easiest thing to miss when planning. The official OpenAI help page is unambiguous:

The Sora web and app experiences were discontinued on April 26, 2026. The Sora API will be discontinued on September 24, 2026.

If Sora is still in your architecture, you have under two months to migrate. It is also a ready-made case study in single-vendor risk.

6. Open weights, and a widely-repeated error

The conclusion first: Wan 2.7 has no open weights. It is API-only.

A large number of articles — including several that surfaced in my first round of searching — describe Wan 2.7 as Apache 2.0 open weights you can download and self-host. First-party sources do not support this. The Wan-Video GitHub organization currently holds five repositories, the newest video model being Wan2.2. The Wan-AI organization on Hugging Face likewise tops out at the Wan2.2 family. Runpod puts it most directly:

You may see third-party posts claiming an Apache 2.0 release for Wan 2.7; that isn't supported by Alibaba's announcement or by anything on Hugging Face.

Since Wan 2.5, Alibaba has kept new generations behind the API; 2.2 is the last with downloadable weights. So the Wan 2.7 on the leaderboard is a hosted API model — do not plan a self-hosted deployment around it.

Here is what you can actually self-host for I2V today:

Cosmos 3 deserves a caveat about positioning: NVIDIA builds it as a world foundation model for physical AI, aimed at synthetic training data for robotics, autonomous driving, and industrial simulation — not at filmmaking. If it feels wrong for creative work, that is because it was never built for it.

7. Pricing: use official pages, because secondary numbers are badly wrong

This was the most valuable part of the verification pass. Secondary review sites quote Veo 3.1 anywhere from $0.03/s to $0.75/s, a 25x spread. The actual figures from Google's official Gemini API pricing:

Runway's official API pricing bills in credits at 1 credit = $0.01: gen4.5 is 12 credits/s ($0.12/s), gen4_turbo is 5 credits/s ($0.05/s), and gemini_omni_flash image-to-video is 10 credits/s plus 1 credit for the first-frame image. A useful cross-check: veo3.1 (audio) on Runway is 40 credits/s = $0.40/s, matching Google's direct price exactly, so the two sources corroborate each other.

Kling's official API docs sit behind a login wall, but the rates from its 2026-02-06 VIDEO 3.0 User Guide are quoted consistently by multiple sources: 6 credits/s at 720p and 8 credits/s at 1080p without audio; 9 and 12 credits/s with native audio; voice control adds 2 credits/s. The dollar conversion depends on the prepaid package, with unit prices between $0.10 and $0.14, so 720p without audio lands around $0.06–$0.084/s.

One detail that gets overlooked: failed Kling API tasks do not consume credits, while the consumer app charges for them. Over heavy iteration that difference adds up.

8. The limits: long clips are still stitched

Veo 3.1 is the only model here that emits synchronized audio in a single generation. Everything else needs a separate TTS or sound-effects pass, adding latency and cost that people routinely leave out of their totals.

The more fundamental limit is duration. The whole DiT family shares one failure mode: attention cost grows quadratically with token count, making long clips exponentially expensive, and long-range temporal coherence is weak to begin with. This is not an engineering detail but an open research problem — ICML 2026 ran an entire workshop on it (From Frames to Stories), and its problem statement is worth quoting directly:

Across minutes of generation, current systems often suffer from identity drift, scene inconsistency, narrative breakdown, and weak responsiveness to user intent.

In practice: most models generate 5 to 10 seconds stably and start drifting past that. Treat long-form video as an engineering problem of stitching shots together, not as something one generation will hand you.

Overall

Work the selection backwards by asking four questions:

1. Do you need native audio? If yes, it is Veo 3.1, because nothing else does it in one pass. If no, do not pay that premium.
2. Can the data leave your infrastructure, and what is your volume? For self-hosting, start with HunyuanVideo 1.5 or LTX-2.3. In the Wan family, target 2.2 — do not be misled by the open-source claims around 2.7.
3. How fine does the control need to be? For specified camera motion, locked characters, or cross-shot consistency, look at multi-reference and motion conditioning support, not at the leaderboard. Runway still has the best control surface here even though its Elo has been overtaken.
4. What is the cost structure? Per second, Veo 3.1 Lite ($0.05/s) to Standard ($0.40/s) is an 8x spread. Confirm your distribution channel actually needs that quality tier.

One closing note: information in this field has a short half-life, and the error rate in secondary coverage is startlingly high. Both the Wan 2.7 licensing claim and the Veo pricing figures above are widely repeated and wrong. Before committing to a purchase, spend a minute on the official pricing page and the official repo — it is cheap, and it saves a lot of trouble.

References

Papers and technical reports

• Image-to-Video Diffusion: From Foundations to Open Frontiers (arXiv:2605.17248)
• HunyuanVideo 1.5 Technical Report (arXiv:2511.18870)
• Scalable Diffusion Models with Transformers (arXiv:2212.09748)
• Classifier-Free Diffusion Guidance (arXiv:2207.12598)
• Stable Video Diffusion (arXiv:2311.15127)
• DynamiCrafter (arXiv:2310.12190)
• From Frames to Stories Workshop @ ICML 2026

Models and repositories

• Introducing Cosmos 3 Edge (NVIDIA)
• nvidia/Cosmos3-Super (Hugging Face)
• Tencent-Hunyuan/HunyuanVideo-1.5 (GitHub)
• Lightricks/LTX-2 (GitHub)
• Lightricks/LTX-2.3 (Hugging Face)
• Wan-Video/Wan2.2 (GitHub)
• Wan-AI (Hugging Face organization)

Official pricing and announcements

• Gemini Developer API pricing
• Runway API Pricing & Costs
• What to know about the Sora discontinuation (OpenAI)
• Kling 3.0 vs O3 API Pricing (EvoLink)
• AI Video API Pricing in 2026 (Apiframe)
• Wan 2.7 on Runpod: open-weight status
• Image-to-Video Leaderboard snapshot (Artificial Analysis / BenchmarkList)

Related posts on this site

• The 3D Modeling Tool Landscape in 2026
• Driving Video Generation Tools with AI Agents: HyperFrames, HeyGen, Runway