feat: RIFE and FILM frame interpolation model support (CORE-29)

[kijai]

Adds pytorch only optimized support for RIFE (MIT) and FILM (Apache 2.0) video frame interpolation models:

https://huggingface.co/Comfy-Org/frame_interpolation/tree/main/frame_interpolation

[coderabbitai]

Note

Reviews paused

It looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the reviews.auto_review.auto_pause_after_reviewed_commits setting.

Use the following commands to manage reviews:

• @coderabbitai resume to resume automatic reviews.
• @coderabbitai review to trigger a single review.

Use the checkboxes below for quick actions:

• ▶️ Resume reviews
• 🔍 Trigger review

📝 Walkthrough

Walkthrough

Adds frame interpolation support via a new extension module comfy_extras/nodes_frame_interpolation.py that registers two nodes (FrameInterpolationModelLoader, FrameInterpolate) and an entrypoint. Introduces two model implementations (FILMNet in comfy_extras/frame_interpolation_models/film_net.py, IFNet in comfy_extras/frame_interpolation_models/ifnet.py) and a FrameInterpolationModel I/O type. Model loader searches models/frame_interpolation, detects FILM vs IFNet weights, normalizes state-dict keys, instantiates and loads the appropriate model, and returns a ModelPatcher. The interpolation node handles padding, optional torch.compile, per-pair multi-timestep inference (with fallback), dtype/device selection, CUDA OOM retry logic, and returns the interpolated sequence. folder_paths.py and builtin extra node imports are updated.

🚥 Pre-merge checks | ✅ 2 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 8.20% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (2 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly summarizes the main change: adding RIFE and FILM frame interpolation model support, which matches the primary purpose of the pull request.
Description check	✅ Passed	The PR description directly relates to the changeset, describing the addition of RIFE and FILM frame interpolation model support with PyTorch optimization.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@comfy_extras/nodes_frame_interpolation.py`:
- Around line 107-110: The padding call on frames using F.pad with
mode="reflect" can fail when pad_h >= H or pad_w >= W (e.g., 16×64 or 32×32);
update the logic around pad_h/pad_w calculation so that before calling F.pad you
choose mode = "reflect" if pad_h < H and pad_w < W, otherwise use mode =
"replicate" (or another non-reflect fallback) and pass that mode into
F.pad(frames, (0, pad_w, 0, pad_h), mode=mode) so the function (and the
variables pad_h, pad_w, H, W and the F.pad call) handles small-frame cases
safely.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: b9519d0a-33fd-4ed3-8b9b-b8a2b44211ad

📥 Commits

Reviewing files that changed from the base of the PR and between 0c63b4f and a859152.

📒 Files selected for processing (4)

• comfy_extras/nodes_frame_interpolation.py
• comfy_extras/rife_model/ifnet.py
• folder_paths.py
• nodes.py

[coderabbitai]

Actionable comments posted: 2

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@comfy_extras/nodes_frame_interpolation.py`:
- Around line 161-167: The multi_fn branch (where multi_fn is used to compute
mids in forward_multi_timestep()) currently asks for all timestep outputs at
once and bypasses the OOM-aware batch-halving fallback; update this branch to
catch CUDA OOM (or memory errors) from multi_fn/forward_multi_timestep and retry
by splitting t_values into smaller chunks (e.g., half the size repeatedly) and
calling multi_fn/forward_multi_timestep per-chunk, copying each chunk into
result as done now, using the same dtype/non_blocking logic and pbar/tqdm
updates; ensure you clear caches between retries (feat_cache or
torch.cuda.empty_cache()) and propagate the original error if even single-step
chunks fail.
- Around line 151-189: The inference loop currently calls inference_model and
inference_model.extract_features while only setting eval(), which still builds
autograd graphs and retains tensors in feat_cache; wrap the loop that iterates
over frame pairs (the try: for i in range(total_pairs): ... finally:) in
torch.inference_mode() so all calls to extract_features, inference_model(...)
and multi_fn(...) run without tracking gradients, preventing unnecessary graph
retention and reducing memory/OOM issues; update any context usage around
feat_cache, multi_fn, and inference_model to execute inside the inference_mode
block.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 43f05f12-4f18-425e-bfd3-3376c4458e9f

📥 Commits

Reviewing files that changed from the base of the PR and between a859152 and 257c531.

📒 Files selected for processing (3)

• comfy_extras/frame_interpolation_models/film_net.py
• comfy_extras/frame_interpolation_models/ifnet.py
• comfy_extras/nodes_frame_interpolation.py

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@comfy_extras/nodes_frame_interpolation.py`:
- Around line 168-180: The OOM fallback in the multi_fn branch incorrectly uses
"continue", which skips writing outputs for the current pair (variables:
multi_fn, result, out_idx, num_interp) and moves to the next pair; instead,
catch model_management.OOM_EXCEPTION, call model_management.soft_empty_cache(),
set multi_fn = None, and do NOT continue so the loop falls through and retries
the same pair using the single-timestep path (i.e., let the subsequent
single-step interpolation logic run for the current i rather than skipping it).

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: e695b624-086d-4151-9510-1abfe96d97d3

📥 Commits

Reviewing files that changed from the base of the PR and between 257c531 and 3cbd1d5.

📒 Files selected for processing (2)

• comfy_extras/frame_interpolation_models/film_net.py
• comfy_extras/nodes_frame_interpolation.py

[coderabbitai]

♻️ Duplicate comments (1)

comfy_extras/nodes_frame_interpolation.py (1)

177-181: ⚠️ Potential issue | 🔴 Critical

OOM fallback skips writing the current pair's original frame.

When multi_fn raises OOM, the continue at line 181 skips the rest of the loop body including line 201-202 which writes images[i + 1] to the result. This causes frame loss—the interpolated frames for pair i are not written (expected, we want to retry), but the original frame images[i + 1] is also skipped (bug).

The intended behavior should fall through to process the current pair with single-timestep calls instead of skipping to the next pair.

Proposed fix

             if multi_fn is not None:
                 # Models with timestep-independent flow can compute it once for all timesteps
                 try:
                     mids = multi_fn(img0_single, img1_single, t_values, cache=feat_cache)
                     result[out_idx:out_idx + num_interp] = mids[:, :, :H, :W].to(out_dtype)
                     out_idx += num_interp
                     pbar.update(num_interp)
                     tqdm_bar.update(num_interp)
                 except model_management.OOM_EXCEPTION:
                     # Fall back to single-timestep calls for this and subsequent pairs
                     model_management.soft_empty_cache()
                     multi_fn = None
-                    continue
-            else:
+                    # Fall through to process current pair with single-timestep path
+
+            if multi_fn is None:
                 j = 0
                 while j < num_interp:

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@comfy_extras/nodes_frame_interpolation.py` around lines 177 - 181, The OOM
exception handler currently sets multi_fn = None then uses "continue", which
skips the remainder of the loop and prevents writing the original frame
(images[i + 1]); remove the "continue" so execution falls through and the loop
will process the current pair using the single-timestep path. Specifically, in
the except model_management.OOM_EXCEPTION block (where
model_management.soft_empty_cache() and multi_fn = None are set), delete the
"continue" and ensure subsequent logic checks multi_fn (now None) and executes
the single-timestep interpolation and writes images[i + 1] to the result.

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Duplicate comments:
In `@comfy_extras/nodes_frame_interpolation.py`:
- Around line 177-181: The OOM exception handler currently sets multi_fn = None
then uses "continue", which skips the remainder of the loop and prevents writing
the original frame (images[i + 1]); remove the "continue" so execution falls
through and the loop will process the current pair using the single-timestep
path. Specifically, in the except model_management.OOM_EXCEPTION block (where
model_management.soft_empty_cache() and multi_fn = None are set), delete the
"continue" and ensure subsequent logic checks multi_fn (now None) and executes
the single-timestep interpolation and writes images[i + 1] to the result.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: e27f8a83-58a9-4a47-ab92-5dfb86d9f203

📥 Commits

Reviewing files that changed from the base of the PR and between 3cbd1d5 and 36a9a60.

📒 Files selected for processing (2)

• comfy_extras/nodes_frame_interpolation.py
• models/frame_interpolation/put_frame_interpolation_models_here

[Kosinkadink]

Tested the code on a windows machine, seems good.

One change I'd request is to remove torch compile from the node. We can worry about that later in terms of compatibility; it's removal will simplify the merge of this PR.

[GogitaTS]

ვინ ხარ ? ორშ, 13 აპრ. 2026, 16:10-ში Jedrzej Kosinski-მა ***@***.***> დაწერა:

…

*Kosinkadink* left a comment (Comfy-Org/ComfyUI#13258) <#13258 (comment)> Tested the code on a windows machine, seems good. One change I'd request is to remove torch compile from the node. We can worry about that later in terms of compatibility; it's removal will simplify the merge of this PR. — Reply to this email directly, view it on GitHub <#13258 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BTI3FEZQDJ2PW7LRE5W2CR34VTKLZAVCNFSM6AAAAACXKZ7MESVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DEMZWGI2DSMBRGQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[GogitaTS]

და მე რატომ მომდის ეს შეყობინებები ? ორშ, 13 აპრ. 2026, 16:11-ში Gogita Tsintsadze-მა ***@***.***> დაწერა:

…

ვინ ხარ ? ორშ, 13 აპრ. 2026, 16:10-ში Jedrzej Kosinski-მა ***@***.***> დაწერა: > *Kosinkadink* left a comment (Comfy-Org/ComfyUI#13258) > <#13258 (comment)> > > Tested the code on a windows machine, seems good. > > One change I'd request is to remove torch compile from the node. We can > worry about that later in terms of compatibility; it's removal will > simplify the merge of this PR. > > — > Reply to this email directly, view it on GitHub > <#13258 (comment)>, > or unsubscribe > <https://github.com/notifications/unsubscribe-auth/BTI3FEZQDJ2PW7LRE5W2CR34VTKLZAVCNFSM6AAAAACXKZ7MESVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DEMZWGI2DSMBRGQ> > . > You are receiving this because you are subscribed to this thread.Message > ID: ***@***.***> >

[Kosinkadink]

@kijai was about to merge, but realized something - do we have any other nodes in core that have a 'model' name for input that aren't a MODEL type? If not, maybe that input can be called interp_model instead. Comparison with LATENT_UPSCALE_MODEL: