Wan 2.2 / LTX-2 OOM Guide: Safe Frames, Batch Size, and Resolution in AI Toolkit

For video LoRA training, OOM usually is not about one bad setting.

It is usually the combination of:

• too many frames
• too large a resolution bucket
• too large a batch
• too expensive a preview sample

That is why video OOM feels inconsistent: one run works, the next run

crashes, even though "nothing important changed."

This guide gives you a practical memory budget for Wan 2.2 and

LTX-2 inside RunComfy AI Toolkit.

---

Quick Fix Checklist (start here)

• For Wan 2.2, start with Batch Size = 1 and 21--41 frames
• For LTX-2, start with Batch Size = 1 and 49 or 81 frames
• In Datasets, lower Num Frames before you touch LR
• In Datasets, drop the highest Resolution bucket first
• In Sample, keep preview videos cheaper than your training budget
• If the log says Bus error / out of shared memory, that is

  not the same as CUDA OOM

---

1) First: know which memory problem you have

CUDA OOM

This guide is for errors like:

CUDA out of memory

OOM during training step ...

Tried to allocate ...

Shared-memory / DataLoader crash

If your log says:

Bus error

out of shared memory

DataLoader worker is killed

That is a different issue related to shared memory (/dev/shm), not GPU

VRAM. See Fix: DataLoader worker Bus error (/dev/shm) error tutorial here->

---

2) The only mental model you really need

For video training, memory pressure rises mainly with:

frames × resolution × batch size

If you increase all three at once, you are very likely building a

borderline run.

---

3) Wan 2.2: safe vs borderline vs high-risk

Safe first run

• Batch Size: 1
• Num Frames: 21 or 41
• Resolution: start with 512
• Keep preview videos conservative

Borderline

• Batch Size: 1
• Num Frames: 81
• Resolution: 480--512

High-risk

• Batch Size ≥ 2 with 81 frames
• High-resolution buckets plus long clips
• Frequent heavy preview generation

Wan rollback order

1. Reduce Num Frames
2. Keep Batch Size = 1
3. Drop highest Resolution
4. Reduce preview cost

---

4) LTX-2: safe vs borderline vs high-risk

Safe first run

• Batch Size: 1
• Num Frames: 49 or 81
• Resolution: 512

Borderline

• Batch Size: 1
• Num Frames: 121
• Resolution: 512

High-risk

• Batch Size ≥ 4 with 121 frames
• Larger buckets before stability is proven
• Heavy preview sampling

LTX rollback order

1. Keep Batch Size = 1
2. Reduce Num Frames (121 → 81 → 49)
3. Reduce Resolution
4. Make preview cheaper

---

5) Why the same config sometimes works and sometimes OOMs

Common reasons:

• Bucket spikes (largest bucket pushes VRAM over limit)
• Preview spikes (training fits, preview pushes it over)
• Borderline memory state

A config that "sometimes works" should be treated as unstable.

---

One-line summary

For Wan 2.2 and LTX-2, video OOM is usually a frames × resolution ×

batch problem.

Start conservative, prove stability, then scale up.