🔨 Forging the Future #2 — The Challenge of Video (Patreon)

Welcome back to the forge.

Last post was about image models.This one? Video.

🎥 Why Wan?

There are several video models out there that I’m sure could perform well. But Wan has consistently been the popular choice in the AI community — and it’s been great for me. Reliable, flexible, and capable of producing strong results.

Because of that, I honestly haven’t felt the need to jump ship and test every alternative.

That said… no model is perfect.

⚙️ The Reality of Limitations

Wan is powerful — but it absolutely has limitations.

Especially when it comes to more complex or explicit scenes.

On its own, Wan often struggles with anatomy accuracy. Certain details simply aren’t well understood by the base model. Without help, you’ll get distorted shapes or abstract “blobs” where precision is required.

That’s where LoRAs come in.

A LoRA is essentially a small, specialized, efficient plugin model used to fine-tune generation toward a very specific outcome. Think of it as giving Wan targeted knowledge it otherwise lacks.

In many cases, I have significantly better results when the original image being animated already contains the correct visual information. If Wan starts with a strong visual foundation, it’s much more likely to animate that element correctly instead of improvising.

🧍 Natural Motion Is Hard

Even with strong source images and LoRAs, motion itself is a massive challenge.

Things like:

How bodies move relative to each other

How parts align from different camera angles

How lighting shifts mid-motion

Facial expressions under movement

Environmental interaction

Every new variable increases complexity.

Many creators train LoRAs using real-life footage to teach Wan how certain motions should look. This helps tremendously — but when you add fantasy characters, non-human anatomy, stylized environments, or exaggerated proportions… the difficulty multiplies.

Sometimes everything just clicks and the output feels incredible.

Other times? The smallest detail — a hand, a background object, a camera tilt — can repeatedly break the scene in ways that require dozens of reruns to fix.

🗡️ A Real Example (End of Act 2)

See the included photo? Remember that scene from the end of Act 2?

That scene perfectly highlights the struggles I’ve dealt with.

I could not find a strong “action” LoRA for a man being stabbed. MANY of the outputs were laughably unrealistic and just plain odd-looking. Limbs bending wrong. Knives floating. Zero weight or impact.

So instead of trying to animate the full motion from scratch, I adapted.

I generated a still image with the knife already in Gideon’s body — roughly correctly placed and angled — and then animated it being pushed further in for dramatic effect.

Was it the “pure” approach? No.

Was it effective? Yes.

Sometimes this process isn’t about forcing the model to do what you want — it’s about understanding its limits and working around them intelligently.

🧪 The Wan 2.2 Transition

Now that I’ve upgraded to Wan 2.2, there’s another layer to this process.

New base model.New behavior.New strengths.New quirks.

And most importantly:

New LoRAs that need to be updated, tested, and optimized.

It’s a whole new can of worms.

But the upside is worth it.

With the upgraded PC and Wan 2.2, I can now generate longer clips with better motion consistency and improved visual quality. That opens creative doors that simply weren’t possible before.

🖥️ The Setup Struggle Is Real

And here’s something most people never see:

It’s honestly crazy how long it takes to get everything fully set up and stable.

PyTorch.Xformers.CUDA.Python versions.SageAttention.ComfyUI.Custom Nodes.Workflows.LoRAs.

One mismatch and something breaks. One wrong compile and performance tanks. One outdated node and your workflow crashes.

I’m just now getting everything to a really solid point on the new PC.

In the next post, I’ll share some examples of what this upgraded machine can generate — and I think you’re going to like what you see.

The forge is hotter than ever.

— Morrak Games