Upscaling

"Upscale" is not one technique. A pixel upscaler, an edit model told to upscale, and the upscaler inside a video model all give different results on different pictures, so Ordinary Animator ships several and lets you (or the assistant) try more than one. Each workflow's description says in one line what it is good and bad at; this page is the longer version.

Because this product makes cartoons, line-art behaviour matters more than photo texture. The question to ask of any upscaler is: do the ink lines stay clean, and does flat colour stay flat?

Upscaling an image

Four ready-made workflows are listed under the Upscale operation: ClearReality, Real-ESRGAN Anime, 4x UltraSharp and Qwen Image Edit 2511.

Pixel upscalers (fast, faithful)

A pixel upscaler is a small model that looks at the picture and outputs a bigger one. There is no prompt and no base model; it cannot add a thing that was not there, so the subject cannot drift. They take seconds and cost very little. All three below are 4x.

Real-ESRGAN Anime (Upscale 4x, Real-ESRGAN Anime)

• Pros: built for drawings. Keeps ink lines clean and flat colour flat.
• Cons: photographs come out smooth and painted. Cannot repair a blurry source, only enlarge it.
• Chains with: anything. Run it first, then a diffusion or edit pass if you want more invented detail.
• Pick this when: the frame is a cartoon, anime or line-art still. Start here for this product.

4x UltraSharp (Upscale 4x, 4x UltraSharp)

• Pros: crisp edges and texture; good on photos, painted backgrounds and rendered frames.
• Cons: over-sharpens. Soft cartoon outlines can grow halos or a "crunchy" edge.
• Chains with: anything; a good first step before a pass that adds texture.
• Pick this when: the frame is a photograph, a painted background or a 3D render.

4x ClearReality (Upscale 4x, ClearReality V1)

• Pros: adds detail without heavy sharpening; a safe general-purpose default.
• Cons: tuned for photographic sources, so it is not the best choice for flat cartoon colour.
• Chains with: anything.
• Pick this when: you want a gentle upscale of a realistic or semi-realistic image.

Edit model told to upscale (invents detail, can drift)

Qwen Image Edit 2511 (Upscale, Qwen Image Edit 2511)

• Pros: re-draws the picture, so it can invent plausible fine detail, rescue a soft, blocky or compressed source, and add texture a pixel upscaler cannot.
• Cons: it re-draws rather than interpolates, so it is the least faithful option. Faces, text and thin lines can drift. It is slow and heavy compared with a pixel upscaler, and its output is capped at about 2 megapixels regardless of the scale you would like.
• Needs: a prompt. The shipped one asks it to enlarge while keeping every shape, pose, colour and line, and not to add or remove anything. For a cartoon, edit it to name what must stay: "keep the black outlines and the flat colour fills."
• Chains with: a pixel upscaler. Run Real-ESRGAN Anime first for a faithful 4x, then (if the result still looks flat) this as a final pass over a smaller crop of it.
• Pick this when: the source is poor and a faithful enlargement just gives you a big poor picture. Always compare against the original before you keep it.
• Release: the workflow uses the 2511 release (the one OA has in its model list). A Comfy Cloud upscale add-on for 2511 exists, but it is not in OA's model list, so it is not used; the shipped workflow works from the prompt alone.

Upscaling a video

LTX spatial and temporal upscalers

These are latent upscalers: they do not work on pixels. They enlarge (spatial) or add frames to (temporal) the internal representation an LTX video is made of, which is why they live inside an LTX generation and not on their own.

• Do they need latents? Yes. They need the LTX video model's own encoder/decoder in the same graph, and an LTX latent to act on. In practice that means they run as the second stage of an LTX generation: generate small and fast, upscale the latent, then decode. This is why OA does not list them as a standalone Upscale workflow. The LTX 2.3 lip-sync workflow already works this way: it runs the 2.3 spatial upscaler as the last stage of its own generation.
• Can a finished video be upscaled on its own? Only by encoding it back into an LTX latent first, which re-processes the video and has not been verified here, so no standalone workflow ships. For a finished clip, use one of the Video to Video upscale workflows below instead.
• Can they be combined with other approaches? Yes, naturally: the spatial upscaler is a second stage, and a pixel upscaler can still be run on the decoded frames afterwards.
• Which release? Use the 2.3 upscalers. The 2.5 download files are gated and are not distributed by OA.
• Pros: cheap high-resolution video, because the heavy work happens at the small size.
• Cons: tied to LTX, so it does nothing for a Wan or MiniMax clip or for a still image.
• Pick this when: you are already generating with LTX and want more resolution from it.

Upscaling a finished clip

Two Video to Video workflows upscale a finished clip frame by frame, both marked experimental: Video to Video (RealESRGAN Anime, Upscale), the cheap baseline that sharpens line art and invents nothing, and Video to Video (SeedVR2 3B, Upscale), a restoration model that vendor comparisons rate best for anime line art, local only. Pick the RealESRGAN one first; compare with SeedVR2 when the take is soft. SUPIR is in the model list but has no workflow yet.

Quick choice

Source	Try first	Then compare with
Cartoon or line-art still	Real-ESRGAN Anime	Qwen Image Edit 2511 with a prompt naming the lines
Photo or rendered frame	ClearReality	4x UltraSharp
Poor, compressed source	Qwen Image Edit 2511	Real-ESRGAN Anime, then judge which is less wrong
Finished video clip	Video to Video (RealESRGAN Anime, Upscale)	Video to Video (SeedVR2 3B, Upscale), local only
LTX generation in progress	The LTX 2.3 spatial upscaler as the second stage	--

When the assistant is asked to upscale something and "try two ways", it picks two from this list by their descriptions and runs both.